JPS59127099A - Improvement in continuous voice recognition - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to speech recognition methods and apparatus, and more particularly to methods and apparatus for recognizing keywords in continuous speech signals in real time.

Various speech recognition systems have been proposed in the past for recognizing isolated utterances by comparing a suitably processed unknown isolated speech signal with one or more child-prepared known keyword representations. I came. The term "keyword" is used herein to mean a group of connected phonemes and sounds, such as a part of a syllable, a word, a word string, a phrase, etc. While many systems have had limited success, one device in particular has been successfully used commercially to recognize isolated keywords. This system is 19
It operates substantially in accordance with the method described in U.S. Pat. This method provided a method for recognizing one of a limited range of keywords, subject to either background noise or silence being measured, and this method has yielded good results. This system relies on the assumption that the period of time giving rise to the unknown audio signal is sufficiently limited and includes only the utterance of a single keyword.

In a continuous audio signal such as an aside, the boundaries between keywords cannot be recognized as wavy lines, but in order to classify the incoming audio data, i.e., phoneme, syllable, word 1.
Various methods have been devised to determine the boundaries of linguistic units such as sentences prior to the start of the keyword recognition process. However, these conventional continuous speech systems have had limited success, in part because no satisfactory segmentation method has been found. Furthermore, there are other substantial problems, such as - comprehensively, only a limited vocabulary can be recognized with a low false alarm rate, and recognition accuracy is highly sensitive to differences in the speech characteristics of different speakers. The system is highly sensitive to distortions in the audio signal being analyzed, such as those common to audio signals transmitted in common telephone communication equipment.

U.S. Patent Nos. 4,227,176 and 4,241,32
No. 9 and No. 4,227,177, respectively, describe commercially acceptable and effective techniques for recognizing keywords in continuous speech in real time. The general methods described in these patents are now commercially available and have been shown experimentally and in practical trials to provide high fidelity and low error rates in speaker-independent situations. It was. However, even these techniques, which are at the current state of the art, have drawbacks in both false alarm rates and speaker-independent characteristics.

The continuous sound Wt fabrication method described in the above-mentioned U.S. patents is as follows:
It is primarily intended for "open vocabulary" situations, where one of several keyboards of continuous speech is recognized or confirmed. The "Open Vocabulary" method is a method in which the device does not understand all of the tasks that are to come. In certain applications, continuous word strings may be recognized, where each individual word element of the continuous word string is identified as a result of the recognition process. A continuous word string, as used herein, refers to a plurality of recognizable elements bounded by silence (a "closed topocytria"). This is related to one of the commercial arrangements described above, for example in connection with the application of isolated words whose boundaries are demonstrably known. However, here the boundaries, ie silence, are unknown and must be determined by the recognizer itself. Additionally, the element being tested is no longer a single word element, but a plurality of elements aligned to form a word string.

In the past, various methods and devices have been proposed for recognizing continuous speech, but little is known about automatically training the devices to generate the necessary parameters to enable accurate speech recognition. I wasn't paying attention.

Additionally, while the methods and apparatus for determining silence and the use of grammatical syntax in prior art devices are generally sufficient for their needs, they still leave much room for improvement.

Therefore, it is a primary object of the present invention to provide a method and apparatus for audio fabrication that is effective in training a device to generate new recognition patterns.

A particular object of the present invention is to effectively recognize silence in unknown audible input signal data, employ grammatical syntax in the recognition process, and respond equally well to different speakers and thus to different speech characteristics. However, it is an object of the present invention to provide such a method and apparatus that is reliable, has a low false alarm rate, and operates in real time.

The present invention relates to a speech analysis method and apparatus for recognizing at least one keyword in an audio signal.

In one particular aspect, the present invention relates to a method for recognizing silence in an incoming audio signal. The method includes generating at least first and second target templates representing alternating displays of silence, comparing an incoming audio signal to the first and second target templates, generating a value representing a result of the comparison, and at least The present invention is characterized in that it is determined whether or not Silent 5 is detected based on this value.

In another aspect, the invention generates a value representative of the likelihood that a current incoming audio signal portion corresponds to a reference pattern representing silence, and advantageously modifies this value according to a syntax-dependent measure; The present invention also relates to a method for recognizing silence in an audio signal, characterized in that it is determined from the effectively modified score whether the signal portion corresponds to silence. Thus - a syntax-dependent measurement indicates the recognition of the immediately preceding part of the speech signal according to the grammatical syntax.As yet another aspect, the present invention provides a method for representing known keywords and The present invention relates to a method of forming a reference pattern adjusted to a reference pattern. This method forms a speaker-independent reference pattern representing a keyword, uses this speaker-independent reference pattern to determine the boundaries of the keyword in the audio signal spoken by the speaker, and determines the boundaries of the keyword in the audio signal spoken by the speaker. The speech analysis device is trained for the speaker using the boundaries determined by the device for the keywords.

The method further includes forming a speaker-independent reference pattern representing a keyword known to the device, determining boundaries of the unknown keyword using the speaker-independent reference pattern, and determining the boundaries of the unknown keyword in advance. The present invention relates to a method for forming a reference pattern representing a previously unknown keyword, characterized in that a speech analysis device is trained using boundaries predetermined by the device to generate statistical data representing the previously unknown keyword.

In yet another aspect, the invention relates to a speech recognition method in which a set of keywords being recognized is described by a grammatical syntax characterized by a plurality of connected decision nodes. This speech recognition method forms a set of numerical scores for recognizing keywords in the speech signal, employs dynamic programming, and uses grammatical syntax to determine which score is an acceptable progression in the recognition process. Decide whether to reduce the number of otherwise acceptable progressions by folding the syntax judgment node, and then reduce the number of otherwise acceptable progressions according to the folded syntax. characterized by the abandonment of

The invention further relates to a device implementing the above-described speech recognition method.

Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.

In certain expedient embodiments described herein, one speech recognition and training involves specially configured electronic equipment that performs certain analog and digital processing of incoming audible audio signals, generally speech, and certain other The data conversion steps and numerical evaluations are accomplished by a system that includes a general purpose digital combiner programmed in accordance with the present invention. The division of tasks between the hardware part and the software part of the system was done in order to obtain a system that can perform speech recognition in real time at a low cost. However, some portions of the tasks being accomplished with the hardware of this particular system could be accomplished satisfactorily in software, and some portions of the tasks being accomplished with the software programming of this particular example may be accomplished with other implementations. In an example, this could be accomplished with special purpose circuitry. This latter KM series describes a single device hardware and software embodiment, if available.

According to one aspect of the present invention, an apparatus is provided for recognizing keywords in a continuous audio signal even when the signal is distorted by, for example, telephone silver. In other words, 1. Especially in FIG.
The voice input signal designated at 10 can be thought of as a voice signal generated by a carbon handset and handset over a telephone line spanning any distance and encompassing any number of switches. Therefore, a representative example of the invention is the recognition of continuous word strings of voice data received over a telephone system, provided by an unknown source (speaker independent system). On the other hand, the input signal may be any audio signal, such as an audio signal derived from a wireless communications link, such as a commercial broadcast station, a private communications link, or an audio input signal of an operator standing in the vicinity of the device.

As can be seen from the above description, the methods and apparatus of the present invention relate to the recognition of audio signals that include a series of sounds, phonemes, or other recognizable symbols. In this specification, any one of "word J1 "element", "series of target patterns", "template pattern", or "element template" will be referred to, but these five terms are general terms. Yes, h is considered equivalent. This is a convenient way to express a recognizable series of sounds or H alternatives that combine to constitute a keyword that can be detected and recognized by the present method and apparatus. These terms should be interpreted broadly and generally to encompass everything from a single phoneme, syllable, or sound to a series of words (in the grammatical sense) as well as a single word.

An analog/digital (A/D) converter 13 receives the incoming analog/4G audio signal data on line 10 and converts the signal width of the data into digital form.

The exemplary A/D converter converts input signal data to a 12-bit binary representation, which is converted 8000 times/
It happens in seconds. In other embodiments, other sampling rates may be employed. For example, if a high quality signal is available, a speed of 16KH2 can be used (A/I)
Converter 13 passes its output through l115 and supplies it to autocorrelator 17. Automva unit 17 processes the digital input signal to generate a short term autocorrelation function 100 times in 11 seconds and provides its output via 11A 19 as shown. Each autocorrelation function has 32 values or channels, and the range is calculated into a 30-bit solution. The autocorrelator will be explained in more detail later in connection with Figure 1a2. Generates a power spectrum.

The spectra are generated with the same number of repetitions as the autocorrelation function, ie, at a rate of 1oo, and each short-term power spectrum has 31 numerical periods with each 16-bit solution. As will be appreciated, each of the 31 periods of a spectrum represents a single perf within a frequency band. The Fourier transform device may also include a Hanning or similar window function to reduce unwanted adjacent band responses.

In the illustrated embodiment, the Fourier transform as well as subsequent processing steps are preferably carried out using a suitably programmed general purpose computer, with separate peripherals to speed up the operations required iteratively according to the method. Preferably, it is performed under the control of a digital computer. The computer used is an FDP manufactured by Digital Equipment Corporation located in Massachusetts.
-11 type. The particular array processor employed is disclosed in U.S. Pat. No. 4,228, assigned to the assignee of this application.
．． No. 498. The program described below in connection with FIG. 3 is set up largely based on the capabilities and characteristics of these available digital processing units.

The short term windowed Howa vector is equalized for frequency response as indicated at 25.

This equalization is thus performed as a function of the peak amplitudes occurring within each frequency band or channel, as will be shown in detail below. The frequency response equalized spectrum on line 26 is generated at a rate of 100 times,
Each spectrum is then evaluated with 16-bit precision and 5
It has 31 numerical periods. To facilitate the final evaluation of the incoming audio data, the windowed spectrum equalized frequency response on 1s26 is subjected to an amplitude transformation as indicated at 35. This imposes a non-linear amplitude transformation on the incoming spectrum /I/. This transformation will be described in more detail below, but at this point it is worth mentioning that it improves the accuracy with which the unknown incoming speech signal can be matched to the target pattern template of the reference speech. In the illustrated embodiment, this transformation is performed on all of the frequency-equalized windowed spectra at some point in time before comparing the spectra to the elemental wandering pattern of the reference speech.

The amplitude transformed and equalized short term spectrum on line 38 is then compared to the element template at 40 as described below. The reference pattern designated 42 represents the elements of the reference pattern in statistical terms with which the transformed and equivalent spectra can be compared. Each time "silent" is detected, a determination is made as to the identity of the word string just received. This is indicated at 44. K like this
The candidate words are selected according to the stringency of the comparison K, and in the illustrated embodiment, the selection process is designed to minimize the possibility of missing or replacing keywords.

Referring to FIG. 1A, the speech recognition system of the present invention employs a controller 45, which includes, for example:
FI) It may be a general purpose digital computer such as F-11. In the illustrated embodiment, controller 45:
The sub-processed audio data is received from the preprocessor 46. The preprocessor will be described in detail in conjunction with FIG.

Preprocessor 46 receives audio analog signals via llA 47 and provides processed data to a control processor or controller via interface line 48.

Generally, the operating speed of the control processor is not fast enough to process incoming data in real time if it is a general purpose processor. As a result, it is advantageous to employ various special purpose hardware to effectively increase the processing speed of element 45. No. 4,22, assigned to assignee of the present invention.
Becmel processing apparatus 48 as described in No. 8,498
a provides significantly increased array processing power by exploiting pipeline effects. In addition, as detailed in connection with FIGS. 4.5 and 6, the one-likelihood function processor 48b can be used in conjunction with a vector processor to further increase the operating speed of the device by a factor of ten. In the preferred embodiment, the control processor 45
In another particular embodiment, which is a digital computer but will be described in connection with FIG. The structure of this processor will be explained in detail later in connection with FIG.

As such, the device for implementing speech recognition exemplified herein has great versatility in its speed and in that it can be implemented in hardware, software, or an advantageous combination of hardware and software. be.

Next, the processor will be explained.

In the device illustrated in FIG. 2, the autocorrelation function with its inherent averaging effect is applied to an analog-to-digital converter that operates on incoming analog audio data, typically an audio signal, provided via an MIO. This is performed on a digital data stream generated by 13. converter 1
3 generates a digital input signal on ll115.

Digital processing functions as well as analog-to-digital conversion are timed under clock oscillator 510 control. The clock oscillator generates a basic timing signal of 256,000 pulses per second, and this signal is passed to the frequency divider 52.
is applied to obtain a second timing signal of s, ooo pulses/second. The low speed timing signal is Ana! It controls the digital converter 13 and the latch register 53. This latch register thus holds the 12-bit result of the last conversion until the next conversion is completed.

The autocorrelation product is generated by a digital multiplier 56 that multiplies the number contained in register 53 by the output of a 32-word shift register 58.0 The autocorrelation product is 1 because register 58 operates in a circular mode and is driven by a fast clock frequency. One rotation of shift register data is performed for each analog to digital conversion. shift register 58
The power for is supplied from the -ff register 53 during one circulation cycle. Digital multiplexer 56
One input is provided directly from the latch register 53 and the other input is provided via the multiplexer 59 from the current output of the 1 shift register. Multiplication is performed at high flux clock frequencies.

In this way, each value obtained from the A/D conversion is multiplied by each of the previous 31 conversion values. As will be apparent to those skilled in the art, the signal thereby generated is equivalent to multiplying the input signal by a signal delayed by 32 different time increments (one being delay is 0). To obtain a zero delay correlation, i.e. to produce a product of the signals, multiplexer 59 multiplies the current value of latch register 53 by itself at the time each value is being introduced into shift register 60. This timing function is indicated at 60.

As will be apparent to those skilled in the art, the product of a single transform and its 31 predecessors does not fairly represent the energy distribution or spectrum for a given sampling interval.

Therefore, the apparatus of FIG. 2 averages these multiple sets of products.

The accumulation step for averaging is connected to an adder 65 and
Provided by 32 word shift registers 63 forming a set of 32 accumulators. That is, each word can be added to the corresponding increment from the digital multiplexer and then recirculated. This circulation loop passes through a gate 67 which is controlled by a divider-by-N 69 driven by a low frequency 4K signal. The divider 69 is a shift register 63
The low frequency clock is divided by 7 actors that determine the number of instantaneous autocorrelation R numbers that are accumulated and thus averaged until the R is read out.

In the illustrated implementation, 80 samples are accumulated before being read. In other words, N for N divider 69 is equal to 80. After the 80 transform samples have been correlated and accumulated, divider 69 triggers computer interrupt circuit 71 via ll1I72. At this point, the contents of the on-schiff register 63 are serially read into computer memory via appropriate interface circuitry 73.

The 32 consecutive words in the register are presented in sequence to the computer via interface 73. As will be apparent to those skilled in the art, this data transfer from one peripheral unit, the autocorrelator preprocessor, to the computer will normally be accomplished by direct memory access methods. 80 samples were averaged with an initial sampling rate of 8000.

It can be seen that 100 averaged correlation functions are provided per second.

While the contents of the shift register are being read from the computer, gate 67 is closed so that each word of the shift register is reset to zero, allowing the accumulation process to resume.

Expressed mathematically, the operation of the device shown in FIG. 2 can be described as follows.

Assuming that the analog-to-digital converter generates a time sequence S (t,), where k = O, To, 2To
,...

To is the sampling interval (1/8 in the illustrated example)
000 seconds)), the exemplary digital correlation circuit of FIG.
Ignoring the start-up ambiguity, one can think of calculating the following correlation function:

a(j, t)=Σs (10 kTo) s (t + (
k j) To) (1) k=.

Here k=o, 1.2. −． 31, t = 80To
.

160'I'o...+80 nTo t... These correlation functions correspond to the correlation outputs on line 19 of FIG.

Referring to FIG. 3, the digital correlator operates to continuously transmit a series of data blocks to the computer at a rate of one correlation function every 10 milliseconds.

This is indicated in FIG. 3 [77]. Each data block represents an autocorrelation function induced into a corresponding subdivision time interval. As mentioned above, the exemplary autocorrelation function is provided to the computer at a rate of 100 32 word functions per second. This analysis interval is referred to below as a "frame".

In the first illustrative embodiment, the processing of the autocorrelation function is
It is carried out on a suitably programmed dedicated digital computer. A flowchart containing the functions provided by the computer program is shown in FIG. However, various of the steps can also be performed by hardware (described below) rather than software, and some of the functions performed by the apparatus of FIG. It should be pointed out that this can also be accomplished with software through modification.

Although the digital correlator of FIG. 2 performs a time-averaging operation of the instantaneously generated autocorrelation function, the average correlation function read out to the computer may interfere with the sequential processing and evaluation of the samples. Including anomalous discontinuity or heterogeneity of species.

Therefore, each block of data, ie each autocorrelation function a(j,t), is first smoothed in time.

This is indicated at 78 in the flowchart of FIG. The preferred smoothing method is the smoothed autocorrelation output as(j
+t) is given by the following formula.

as(j, t)=0oa(j, t,)+
a, a(j, t-T)+02a(j, t-2T)(
2) Here a(j, i) is the unsmoothed input autocorrelation R number defined in equation (1), as(j, t) is the smoothed autocorrelation output, and J represents the delay time. , t represents real time and T represents the time interval (frame) between successively generated autocorrelation functions, which in the preferred embodiment is equal to 0.01 seconds. The weighting is the function 00 +a,,
C+ is preferably 1/4 in the illustrative example.
, 1/2, and 1/4, but other values may also be selected. For example, a smoothing function that approximates a Gaussian impulse response with a cutoff frequency of 20H2 could be implemented in computer software. However, experiments have shown that satisfactory results can be obtained with the easy-to-implement smoothing function exemplified by equation (2)K.

As mentioned above, the smoothing function is applied separately for the delay J. To increase processing speed, the conversion of the autocorrelation function to the frequency domain is performed in the illustrated embodiment with 8-bit arithmetic. At the high end of the bandpass near 31dlz, the spectral power density is reduced to a level insufficient for resolution in 8-bit quantities. Therefore, the frequency response of the system is ramped at a rate of rise of 6 db/octave. This is indicated by 79. This high frequency enhancement is accomplished by taking the second derivative of the autocorrelation function with respect to that variable, ie the time delay, K.

The differential operation is as shown in the following equation.

b(j+t)=-a(j+1.t,)+2a(j,
t)-a(j-1, t,) (To find the differential value with respect to atj=0, the autocorrelation function is symmetric with respect to 0, so a(j, t)-a(+j, t) Also, since there is no data for (32) K,
It is assumed that the differential value at j=31 is the same as the differential value at j=30.

The next step after high-frequency emphasis in the analysis procedure is to find the peak absolute value of the autocorrelation and calculate the signal power in the current frame interval from K as shown in the 70-chart in Figure 3. .

The estimated power p(t) is as follows.

p(t)−〒”l b(1,i)l
(To prepare an autocorrelation function for a 418-bit spectral analysis, the smoothed autocorrelation function is block standardized (at 80) with respect to p(t) and the top 8 bits of each standard value are input to the spectral analysis hardware. is input.

Therefore, the standardized and smoothed autocorrelation function is:

c(j, t) = 127 b(j, t)/p(
t) 俤) A cosine Fourier transform is then applied to each time-smoothed, frequency-enhanced, standardized correlation function as indicated at 81;
Generate a point power spectrum. The matrix of cosine values is given by: That is, g(i, j) ” 126 g(i)(cos (
2πi/8000)f(j)).

j=0,1,2+...,31 (
6) Here, 5(11J) is f(j
) Hz, the spectral energy of the band centered at g(
i)=1/2(1+cos 2πi/63) is the (Hamming) window function envelope for sidelobe reduction, and f(j)=30+1000(0,0552j+0.43
8) 110.6311z, (7) j=0.1,2.
..., 31 These are the analysis frequencies equally spaced on the tonic pitch so-called "mel" curve. As can be seen, this is the dominant pitch (in mel scale) for the frequency of the bandwidth of a typical communications channel of approximately 3000-5000] IZ.
The zero spectral analysis corresponding to the frequency axis spacing is −31
Assuming that the autocorrelation is symmetric with respect to OK, only positive values of J are required since we need to add the delay from to +31. However, the delay O term is reduced to 2
To avoid calculating degrees, the 1 cosine matrix is adjusted as follows.

a(0,j)=126/2=63. for the rest
(8) Thus, the calculated power spectrum is given by the following equation.

where the jth result corresponds to frequency f (j ) 0. As is also clear, each point or value in each spectrum represents a corresponding frequency band.

Although this Fourier transform can be performed entirely within conventional computer hardware, the process could be considerably speeded up by utilizing external hardware multiplexers or fast Fourier transform (FFT) peripherals. However, the structure and operation of this type of module is well known in the art and will not be described in detail here. Advantageously, the hard fair fast Fourier transform peripheral incorporates a frequency smoothing function, in which case each spectrum is given the preferred (Hamming) window weighting described above by the function g(1).
The frequency is smoothed according to This corresponds to 83 of block 85, which corresponds to the implementation of the Fourier transform by hardware.
It will be carried out in

If there is considerable background noise, an estimate of the background power spectrum can be obtained at this stage by S
'(Jlt).

The frame(s) selected to represent noise must not contain any audio signal.

The optimal rule for selecting the noise frame spacing will vary depending on the application. If a speaker engages in intercommunication with a machine controlled, for example, by a whooping recognition device, it is convenient, for example, to arbitrarily select a frame at an interval immediately after the machine has finished speaking with its voice response unit. . If there are fewer constraints, the noise frame can be found by selecting the lowest amplitude frame of the open audio input in the past 1-2 seconds. It is clear that the use of a minimum amplitude "silent" pattern, in fact two alternating "silent" patterns, provides advantageous device operation, as will be explained in detail below.

Once the successive smoothed power spectra are received from the fast seven-tier transform peripheral 85, the peak power spectral envelope (which is −7 general KJ%) for the spectrum from one peripheral 85 is determined as described below. , by changing the output of the fast Fourier transform device accordingly, thereby equalizing the communication channel. Each newly generated peak amplitude corresponding to and modified by the incoming nitriding power spectrum s'(j, t) (where j is assigned to multiple frequencies of the spectrum) is one for each spectral channel. or the result of a fast attack, slow decay, or peak detection function on the band. The wintort power spectrum is normalized with respect to each period of the corresponding peak amplitude spectrum. This is indicated at 87.

In the illustrated embodiment, the "old" peak M width spectrum p(j+tT), determined before receiving the new window toss vector, is the newly arrived spectrum.
(jy t) and frequency bands are compared in a frequency band-to-frequency manner. A new characteristic spectrum p(j, t) is then generated according to the following rules. The power amplitude of each band of the "old" peak amplitude spectrum is in this example a fixed fraction, e.g. 1023
/1024.

This corresponds to the slow decay part of the peak detection function.

If the power amplitude of the frequency band J of the arriving spectrum (jy t) is greater than the power amplitude of the corresponding frequency band of the decay peak amplitude Sve)y, then 1, then the decay peak amplitude spectrum value for that (or those) frequency bands is , is replaced by the spectral value of the corresponding band of the incoming window toss vector. This corresponds to the fast attack part of the peak detection function. Mathematically, the peak detection function can be expressed as .That is, p(j,
t)=max p(j, t-T)(1-E)・p(
t)s(j, t,),' (+(1j=0,1...
・, 31 where %J is assigned to each of the frequency bands and p(
Jlt) is the resulting peak spectrum and p(j.

t, −T) is the old −1 or previous peak spectrum, s′(jet) is the newly arrived partially processed power spectrum, and p(t) is the power estimate at time t. where 1 is the decay parameter.

According to the equation +1 (IK), the peak spectrum typically collapses at a rate of 1, 1-1 in the absence of higher value spectral input.Normally, I is equal to 1/1024 and the signal channel or audio Cases in which rapid changes in properties are not expected1
It would be undesirable to allow the peak spectrum to decay. To limit the silent frames, the same method adopted to select the background noise frames is adopted. The past 128 7th amplitudes (square root of p(t)) are examined and the minimum value is found. If the amplitude of the current frame is less than four times this minimum value, the current frame is determined to be silent and the value ``0'' is substituted for E instead of the value 1/1024.

After the peak spectrum is generated, the resulting peak amplitude spectrum p(j,t) is calculated by averaging each frequency band peak value with the peak value corresponding to the momentarily adjacent frequency of the newly generated peak spectrum. is smoothed (89). Thus, the total frequency bandwidth contributing to the average value is approximately equal to the typical frequency spacing between the formant frequencies. As will be apparent to those familiar with the art of speech recognition, this spacing is approximately 1000 Hz. Averaging in this particular way preserves useful information in the spectrum, ie local fluctuations representing formant resonances, while suppressing the overall enhancement of the frequency spectrum. In a preferred embodiment, the peak spectrum is smoothed in frequency by a moving average function covering seven adjacent frequency bands. The average function is p(k, t) at the end of the bus band at the following dotoku
is O for k less than 0 and k greater than 31
becomes.

The normalization envelope h(j) takes into account the number of valid data elements actually added. Thus, bld=7
/4, h(t)=715, h(2)==7/6,
h(3)=L...

h(goods))2i, hto)) =7/6. kl(3
0)=715. Then, h(31)=7/4. The resulting smoothed peak amplitude spectrum sl,t) is then used to standardize and frequency equalize the just-received pan spectrum, which corresponds to each of the incoming smoothed spectra S'(J+t). This is done by dividing the amplitude values of a frequency band by the corresponding frequency band values of the smoothed peak spectrum e(j,t). Mathematically, this can be expressed as follows.

8n(j, t)=s(j, t)/e(j,
t) 32767 (Ll here, S n (
f+t) is the N-peak normalized smoothed power spectrum and j is assigned for each frequency band. This step is indicated at 91. Here we obtain a series of frequency-equalized and standardized short-term power spectra that emphasize changes in the frequency content of the incoming speech signal and suppress typical long-term frequency enhancements or distortions. It is what was done. This frequency compensation method differs from conventional frequency compensation systems, where the compensation criterion is the average power level of either the entire signal or each frequency hand, which is transmitted over a communications link that produces frequency distortion, such as a telephone line. Although the successive spectra have been variously processed and equalized, the data representing the incoming audio signal is still very useful in the recognition of audio signals occurring at a rate of 100/s. Be sure to point out what is included.

The standardized and frequency equalized spectrum as indicated at 91 undergoes an amplitude transformation as indicated at 91. This is done by performing a non-linear scaling operation on the spectral amplitude values.

5n(j, t) (from Equation 12) (from Equation 12). ), non-linearly scaled spectrum x(j, t
) is the mean value of t) defined by the following linear fractional function: 0, where J indicates the frequency band of the power spectrum.

The 31 periods of the spectrum are replaced by the logarithm of A as follows. That is, x(31,t)=16
log, person O! 9This scale function (
Equation 13) exerts a flexible threshold and gradual saturation effect on spectral intensities that deviate significantly from the short-term average value. Mathematically stated, it is approximately linear for intensities near the average, approximately logarithmic for intensities away from the average, and substantially constant for extreme intensity values. For logarithmic scale, the function "(,l tt
) is symmetric with respect to 0 and exhibits smooth and saturated behavior that exhibits a steep rate function that stimulates the auditory nerve. In fact, the entire recognition system performs considerably better with this particular non-linear scale function than with either case.

In this way, a series of short-term power specs # x (
j, t) (here, t=o, o 1 , 0.02
.

0.03, 0.04 seconds, j=0. ..., 30 (corresponding to the frequency bands of the generated power spectrum) are generated. 32 wads are prepared for each spectrum, A (Equation 1
5), that is, the average value of the spectral values is stored as 32 words. This amplitude-transformed short-term power spectrum, referred to below as a "frame," is, in the illustrated embodiment, as indicated at 95:
It is stored in a first-in, first-out circular memory with a storage capacity for 256 32-word spectra. Thus, in the illustrated embodiment, 2.56 seconds of audio input signal is available for analysis.

This storage capacity, if necessary, provides a protean recognition system that can select spectra at different real times for analysis and evaluation, and thus move forward or backward in time as needed for analysis. Thus, the frame χ for the last 2.56 seconds is stored in circular memory and available when needed. In the illustrated embodiment, each frame in operation is 2.56
Seconds are memorized. Thus, the frame that entered the circular memory at time t1 will be 2.56 seconds later\time t+2.5.
When a new frame corresponding to 6 seconds is stored, it is lost or shifted from memory.

The frames passed through the circular memory are compared against a known range of words, preferably in real time, to determine and identify human data in groups of words called word strings. Each search word is represented by a template pattern that is formed into a plurality of non-overlapping multi-frame (preferably three-frame) design set patterns and statistically represents a plurality of processing power spectra. These patterns are preferably selected to best represent the meaningful acoustic events of the word word and are stored 94 in sequence.

The spectra that form the design set pattern are:
To process the continuous unknown speech input on m10, illustrated in Fj4, the system described above is used to generate words spoken in various situations.

In this way, each word word, generally has a plurality of series of design set words associated with it, p(1)+yp
(1) 2 t..., each pattern giving one indication for its first keyword in the region of the short-term spectrum. The collection of design set patterns for each keyword forms a statistical basis for generating target patterns.

In an exemplary embodiment of the invention, the design set patterns yp(i),l can be thought of as 96-element arrays, each comprising three selected frames arranged in series. The frames forming the pattern are at least 30 to avoid unnecessary correlations due to smoothing with respect to time FC.
Should be milliseconds apart. In the embodiment of the present invention, other sampling methods can be implemented to select frames; however, a preferred method is to select frames at fixed sentence durations, preferably 30 milliseconds apart, and The method is to space non-overlapping design set patterns during time intervals that define keywords. That is, the first design set pattern p, corresponds to the part near the opening point of the keyword, the second pattern p, corresponds to the part after time, and so on, and the pattern pl
r p21... forms a statistical reference voltage or word template for the series of target patterns;
Incoming audio data is matched to this. The target pattern is tl ” 1 r・−・ is p (
i) Consists of statistical data generated from the corresponding p (i) j by assuming that j consists of independent Laplacian variables. This assumption allows likelihood statistics to be generated between the incoming data and the target pattern, which will be described below. Thus, the target pattern consists of an array containing the mean, standard deviation, and area standardization ratio for the corresponding design set pattern array entries collected as a function. More accurate likelihood statistical data will be explained later.

As is clear to those skilled in the art, almost every word has more than one contextual and/or regional pronunciation, and therefore more than one "spelling" of the Design Set pattern. I'm doing it. Thus, a word word with the patterned spelling p1r pz... described above is in practice generally p(')1 + p(i)2
r..., i=1.2. ...It is expressed as 7M.

Here each p(i)j is a possible alternative description for the third class of design set patterns, and there are a total of M different spellings for each word.

Therefore, the target pattern tl r t2+...
, tl are, in the most general sense, each
represents multiple alternative statistical spellings for the second group or class of design set patterns. Thus, in the illustrated embodiment, the term "target pattern" is used in a very general sense, such that each target pattern has two or more permissible alternative "statistical spellings." It is possible.

Preliminary processing of the incoming unknown audio signal and the audio signal forming the reference nodules is now complete.The processing of the stored spectra will now be described.

U.S. Patent Nos. 4,241,329 and 4,227,17
A deeper study of keyword recognition methods that link phonetic patterns to detected words, described in No. 6 and No. 4,227,177, shows that it is a special case of more general and perhaps submerged methods. I understand.

Referring to FIG. 4, the search for word recognition can be expressed as a problem of finding a suitable path in an abstract state space. In this figure, each day has a dwell (
(deferred) represents possible states, also referred to as time positions or registers, through which the decision process can pass. The spaces between the vertical dashed lines 120, 122 each represent hypothetical states that the decision process may pass through in determining whether a pattern matches or does not match the current phoneme. This space is
It is divided into a mandatory dwell time portion 124 and an optional dwell time portion 126. The required dwell time portion is
Represents the minimum duration of the current phoneme or pattern.

Each day within the optional or required dwell time portion represents a frame time of the series of frames formed and corresponds to an interval of 0.01 seconds from frame to frame. Thus, each day represents the hypothetical current phoneme position in one word spelling, and the number of frames (in 0,01 seconds) assumes the time elapsed since the onset of the current phoneme and its speech or target It corresponds to the number of previous circles in the pattern and represents the current continuation of the pattern. After one pattern (phoneme) begins and the minimum dwell time interval has elapsed, there are several possible paths to proceed to the first node or location (circle) 128 of the next target pattern (phoneme). This depends on when the decision is made to move on to the next pattern (phoneme) of spelling. These decision possibilities are represented in this figure by several arrows pointing toward circle 128.

Although the starting point of the next pattern (phoneme) is represented by circle 128, this transformation to the next pattern can occur from any node or position during any dwell time of the current pattern (phoneme), or from any required will be done from the last node of the dwell time interval.

U.S. Patent Nos. 4,241,329 and 4,227,17
The keyword recognition method described in No. 6 and No. 4,227,177 transforms at the first node such that the accuracy score for the next pattern (phoneme) is better than the accuracy score for the current pattern (phoneme) IC. Do the following. That is,
A frame is made at a point that matches the next phoneme or pattern better than the current phoneme or pattern. The total word score, on the other hand, is the average pattern (phoneme) score per frame (ie, per node included in the path). The same fixed definition of "total score" applied to word scores up to the current node can be used to determine when a transformation should be made. That is, it can be used to determine whether the conversion to the next pattern should occur at the first opportunity, eg, corresponding to conversion instruction +l 1130, or at a later point in time, eg, corresponding to conversion instruction line 132. Optimally, one would choose the path in the next pattern (phoneme) that has the best average score per node. US patent jI4
, No. 241,329, No. 4,227.176, and No. 4,227,177. Since we don't test on the right path, we will make a near-optimal decision as measured by the average score/section.

Therefore, the present invention adopts the average score/clause method for keyword recognition. The problem arises in connection with word string recognition, which will be discussed in more detail later, and involves either normalizing all partial word scores by the number of nodes involved (which is computationally inefficient) or Bias should make explicit standardization unnecessary. The natural bias value to be used in the closed-topocylibrary task is the unstandardized score for the best word finishing at the current analysis time. Therefore, the cumulative score at all nodes will always be the sum of the same number of basic pattern scores. Furthermore, the score is converted by this bias value to the score □·1 of the best word string ending at the current analysis node.

The average score/node determination method is described in U.S. Patent No. 4.22.
This can be efficiently implemented using dynamic programming techniques in the vector processor described in US Pat. No. 8,498. When programmed in this manner, the processing speed is reduced even though more hypothesis testing is required. Much faster than the standard keyword recognition methods described in the issue.

Generally speaking, to recognize a word string, the program stores the name of the hypothetically best possible word that ends at each analysis node.

It also remembers the node (time) at which this best word began. Next, pack tracing from the end of the utterance,
The best word string is found by noting the name of the stored word and finding the next aforementioned word at the indicated starting point of the current word.

Including silent as a candy word makes it unnecessary to determine the number of words contained in the word string.The backtracking operation to find the 0 string is performed when the silent word has the best score. and terminates the next time the previous silent is detected. Thus, a string is found every time the speaker pauses for breath.

The word string classification method described herein is a more efficient method of extraction than individual keyword detection. Word string scores are advantageous over simple word spotting methods because they force all sounds in an utterance to be included in a word string. The latter method often detects misspelled words in long words.

An advantage is that no timing pattern is required for the word string case. This is because the word concatenator outputs the word start time for each word end hypothesis. The simplest string concatenator is: Assume that these word start times are correct. Silent detection assumes that the word string has just ended and that the start of the last word (which may also be silent) is the end of the previous word (which may also be silent). Since there are no context-dependent transformations between word pairs, it is preferable to allow the device to search the neighborhood of the start of each word to find the best end of the preceding word. Dew.

Next, WPaFC methods and apparatus, including hardware and software implementations, are described.

Third. To explain with reference to the figure, first, the spectrum or frame stored at 95 representing the incoming continuous audio data is stored in the stored target pattern template (96) representing the keyword of the word search according to the following method.
compared to

For each 10 ms frame, the stored reference pattern and the butter/ for comparison are the current spectrum vector s(j,, t), the spectrum s from three frames ago
(J, t, -o, o3), and the spectrum sQ, t-0,06) from 6 frames ago are shielded to form the following 9
97 by forming a 6-element pattern. As mentioned above, the stored reference pattern is the average value of the previously collected 96-element patterns belonging to the various speech pattern classes to be recognized, the standard Consists of deviation and area standardization factor. The comparison is performed by a probabilistic model of fDix(j,t) that predicts that the input speech belongs to a particular class.

For probabilistic models, Gaussian distributions can be used (e.g., U.S. Pat. No. 4,241,329, cited above, 4.2).
27,176 and 4,227,177),
Laplace distribution, i.e. 1) (X) = (1/v'2 s') exp - (J
21x-m)/s') (where m is the statistical mean and S is the standard deviation of the variable xf) requires less calculation and -
It has been found that it performs much like the Gaussian distribution in the speaker-independent isolated word recognition method described, for example, in US Pat. No. 4,038,503. The degree of similarity L (xlk) between the unknown input pattern (turn I and the th memorized reference pattern) is proportional to the logarithm of the probability, and is expressed as 1
Calculated as 00.

Kokote, Ak=工'Z In 8'1k21=1 Since the likelihood scores L of a series of patterns are combined to form the likelihood score of a spoken word or phrase, the score L(XIX) for each frame is − Adjusted by subtracting the best (minimum) score of all reference patterns for that frame. That is, L'(xlk)
= L (xlk) −min L (xi i)
(II Therefore, the best-fitting pattern for each frame will have a score of 0. The adjusted scores for the hypothesized sequence of patterns are accumulated for each frame and support the indicated sequence of sequences. A sequence score can be obtained that is directly related to the probability that a decision is the correct one.

Comparison of the unknown input vector pattern to the stored known pattern is accomplished by calculating the following function for the kth pattern: That is, 6− q2, Σ sik lXi uik I+
ck a fight 1=1 Here, Sik is equal to 1/s'1kK.

In a calculation performed in normal software, the following instructions would be executed to calculate the algebraic function 5lx-ul (Equation 19).

1. Calculate Zu U. Z Test the sign of Zu U. If Ax-u is negative, negate it to form an absolute value. 4. Multiply by s & add the result to the accumulator. 20-Word Glossary In a typical speech recognition system with a busy system, there would be approximately 222 different criteria (turns).
2=10560 steps, which must be performed within 10 ms to keep up with the real-time spectral frame rate. Therefore, the processor calculates the likelihood function J times, so K is approximately 1100
Must be able to execute instructions at 10,000 per second. Taking into account the requisite speed, a dedicated likelihood R-number hardware module 200 (FIG. 4) is employed that is compatible with the vector processor system disclosed in U.S. Pat. No. 4,228,498. .

In this specialized hard square, the above-mentioned five steps are executed simultaneously with two sets of variables 5sXsu, so ten instructions are actually executed in the time required to execute one instruction.

Since the basic vector processor operates at a speed of 8 million (instructions)/second, the effective calculation speed for the likelihood function is
If 200 dedicated hardware are adopted, the cost will be approximately 80 million (
command) will be 7 seconds.

Referring to FIG. 5, a node 1-ware module 200
employs a combination of hardware piping and parallel processing to enable simultaneous execution of ten steps. Two identical parts 202, 204 each perform five arithmetic steps on independent input data variables, and the results are combined by an adder 206 connected to its output. The accumulation of the summed value from adder 206 is the summation of 1 to 96 in equation (19), and this value is calculated by the arithmetic unit of the standard vector processor described in U.S. Pat. No. 4,288,498. It is processed.

In operation, the pipe fin binding register holds intermediate data for the following processing stages.

1. Input variables (clock-operated registers 208, 210
．． 212.214.216.218) Absolute value of Lx-u (clock operated register 220.222) 5 Multiplier output (clock operated register 224.22)
6) Once the input data is held in the clocked registers 208-218, the magnitude of X-u is determined by the subtraction and absolute value circuit. Referring to FIG. 6, the subtraction/absolute value circuits 22g, 230 select the first and second subtracters (one calculates x-u, the other calculates u-X) and a positive result, respectively. A multiplexer 236 is provided for this purpose.

Input variables X and U on 1123B, 240 coming out of registers 208.210 are -128 to +127, respectively.
is an 8-bit number. Since the difference output of an 8-bit subtracter may overflow to 9 bits (for example, 127
-(-128)=255), extra circuitry is employed to handle overflow conditions. The condition is determined by overflow detector 235. Thus, the inputs are the sign of "I" (on line 235a), the sign of "u" (on line 235a), and the sign of "u" (on line 235a).
235b) and the sign of rx-uJ (1s235c
above).

Referring now to FIG. 7, one overflow detector, in this illustrative embodiment, consists of three ω gates 268.
270 and an OR gate 272. The quintessence table of FIG. 8 represents the overflow condition as a function of input.

The over 7 days-condition is handled by making four selections in multiplexer 236, which is the circuit that selects the positive subtractor output. The selection is +1l12
42 and 244 on a binary level. $24
The level above 2 represents the sign of x-u. If the code on H.244 is 1, it indicates an overflow.

Thus, the choices are as follows.

Line 224 Line 224 0 0 Select the output of subtracter 232 1 0
The output selection multiplexer of subtractor 234 is controlled in this manner to act like an 8-pole 4-position switch. The shifting operation is performed by connecting the subtracted output by the combiner to the appropriate multiplexer. The shift has the effect of arithmetically dividing by two.

If over 70- occurs during subtraction, the output of the multiplexer will be the output of the subtracter divided by four. It is therefore necessary to recall this condition later in the calculation so that the final result can be multiplied by 2 to recover the correct scale factor. This restoration is done in the multiplexer after the last pipe processing register. Therefore, pipe twin processing registers 220, 22
An extra bit is provided at 2% 224.226 to control the second multiplexer 248.250.

The latter multiplexers each have an 8x8 bit multiplier 252 in case of an overflow bit (equal to 1).
．． The multiplication product of 254 is shifted up by 1 bit and multiplied by 2. Multiplication operations can be performed on standard integrated circuit devices, such as the TRY MPY-8-HJ, which accepts 8-bit numbers and outputs the product.

Thus, the multiplier 252.254 produces a product of 100 and 1x-ul on each clock pulse (the 100 value is correctly timed by the extra data register 256.258).
. The output of multiplier 252.254 is output to register 224.2.
26, via line 260.262,
It is output to the remaining circuits via adder 206.

The same dedicated hardware module can also be employed to compute two-vector inner products such as those required in matrix multiplication. This is the subtraction/absolute value circuit 22
8. Gate circuit 264 allowing bypass at 230
．． 266. In this mode of operation, the data inputs and inputs are applied directly to the multiplier processing registers 220, 222, 0th order, 7-)' level pattern alignment.

K performs dynamic programming (101) to optimize the correspondence between unknown input speech and each word template.
is preferably adopted. Each word template is
In addition to the set of reference pattern statistical data described above, it may also include the minimum and maximum dwell times associated with each reference pattern.

According to the dynamic programming method, one storage register is provided for each word. The number of registers is equal to the sum of the maximum dwell times of the reference patterns that make up the word. That is, it is proportional to the longest allowed word duration. These registers correspond to the circles in Figure 4, with one register for each day. For each frame of input audio, all registers are read from and written to 0, each register is written to 1, as detailed below.
It includes an accumulated likelihood score corresponding to the assumption that the indicated word word is being spoken and that the current position in that word corresponds to a particular reference pattern and dwell time for that register. All registers are initialized to contain low likelihood scores, indicating that none of the above assumptions are acceptably likely to occur initially.

The rules for updating the registers are as follows: 0 The first register of each '7-word template (i.e. the register corresponding to the assumption that the word has just begun to be uttered)
is (a) the likelihood score of the current frame with respect to the first reference pattern for that word, and (b) the best score of all registers of all word work words (i.e., if a word was completed on the previous frame). (cumulative likelihood score) for the hypothesis that

The second register of the word template contains (a) the likelihood score of the current frame for the first reference pattern of the word, and (b) the contents of the first register from the previous frame. Thus, the second register contains the score of the hypothesis that the indicated word is being uttered and that it began in the previous frame.

During the process of updating these registers corresponding to the dwell time (arbitrary dwell period) between the minimum and maximum durations,
The best accumulated likelihood score (
A separate memory register is employed to store the contents of the register. This best score found in the previous frame time is used to calculate the next contents of the first register corresponding to the required dwell time of the next target pattern or template for that word. Thus the current content of the first register of the first-order reference pattern is 1 and adds its best score (of the preceding target pattern) to the likelihood score of the current input frame with respect to said next reference or target pattern. It is caused by

In FIG. 4, the multiple arrows leading to the first register 128 of the required dwell interval of the reference pattern indicate that the conversion from an optional dwell time register or state to a required dwell time register or state occurs at any time during the optional dwell time interval. It is meant to indicate the point in time or originating from the last register of the required dwell time interval. Thus, based on the current information, the best adaptation between the word template and the input pattern is that when the next pattern is just starting, the previous pattern contains the best score of any previous dwell period. has a duration corresponding to the last register (register 300 in the illustrated embodiment) of the required time interval.

According to the theory of dynamic programming, there is no need to store previously accumulated scores corresponding to all possible dwell times. This is because, according to this theory, dwell time transformation points that produce low scores will continue to produce low scores in all future processing stages.

The analysis proceeds in the manner described above using all registers of all reference patterns of all word templates.

The last register(s) of the last pattern of each word template contains the score of the hypothesis that the word has just ended.

During likelihood score accumulation, a series of duration counts are maintained to determine the duration of the best word ending in each frame time. The counting starts at "11" in the first register of the first template pattern of the word.For each second and subsequent register of the template pattern, the count values associated with the various registers are incremented by "1". Ru. However, for each register corresponding to the start of a reference pattern (other than the first reference pattern of one word), i.e., the first register 128 of the required dwell time interval, has the best likelihood score in the previous frame time. The count of the optional dwell time register (or the last required dwell time register) of the preceding reference pattern is incremented to form the duration count for the register.

To provide a backtracking,mechanism that will be described in more detail later,,for each frame time, information about the best-scoring word,ending at that time and its duration is transferred to a circular,buffer memory. When a series of words ends, the stored word durations start from the end of the last "best" word and work backwards in duration until "the best preceding word ends just before the last word j. etc., allows backtracing until all words in a word string have been identified.

A string of successively uttered words is bounded by silences. ``Silent'' therefore acts as a control word that delimits the range of search words J that the shift system responds to and recognizes. As mentioned above, it is not uncommon for a device to detect the lowest amplitude signal during a period of time and indicate it as "silent."

However, according to the invention, one of the word templates corresponds to silent or background noise. If the silent word has the best likelihood score, it is presumed that the series of words has just ended and a new series of words has begun. A flag register is tested to see if a non-silent word has had the best score since the last initialization of the recognition process. [If at least l words from siren H onwards have the best score (103), the word string in the circular buffer is backtraced (105) and the resulting recognized message is sent to the display or other control device. communicated. The circular buffer is then cleared to prevent repeated transmission of the message, and the flag register is cleared.

Thus, the device returns the next word string wty
It is initialized to lI4 (107).

Advantageously, in a preferred embodiment of the invention:
One or more "silent" spellings can be used, just like other "keyword" spellings. That is, the device is not only limited to detecting silences when matching a squiggly set of criteria, i.e., when matching a discursive target pattern, but also detecting a dynamically changing target pattern. Alternatively, templates can be employed to further improve the device's "silent" detection capabilities. In this way, dynamic changes can be made, as described above, by periodically testing the preceding 1 or 2 seconds of audio and selecting, for example, the representative pattern with the lowest amplitude during the last few seconds. A "dynamic" silent model can be determined, a previous dynamic silent model can be updated, or a new "dynamic" silent model can be formed according to the training method described below. In this way, a "silent" can be defined by two or more "spellings" of the Tarkerat pattern, and the possibility of improving the accurate detection of the silent is improved.

Next, I will explain the training for the 1 standard hatan.

In order to obtain the sample mean u and parity S' for the construction of the reference pattern, a large number of utterances of each word word are fed into the speech recognition system, and all statistical data of the corresponding preprocessed spectral frame is determined. It will be done. An important and consequential action of the device (i) is the selection of which human scum vector frame should correspond to which target or reference pattern.

In the absence of sufficient information, such as the important acoustic phonemes chosen by a human for a word, the time interval between the start and end of a spoken word can be divided into a large number of uniformly spaced Sweat divided into intervals. Each of these sub-intervals is associated with a 1-only reference pattern. One or more 37-lem turns are formed starting at each interval and sorted according to the reference pattern associated with that interval. Subsequent instances of the same word word are similarly divided into an equal number of uniformly spaced intervals.

The average values and variances of the elements of the three frame patterns extracted from the corresponding ordinal intervals are accumulated over all available columns of a single word to form a set of reference patterns for that word. The number of intervals (number of reference patterns) should be approximately 2 or 3 per linguistic phoneme unit included in the word word.

For best results, the start and end points of each word are marked by a procedure that includes human examination of recorded audio waveforms and spectral frames. To carry out this procedure automatically, K needs to speak one word at a time, one word at a time, so that the device finds the word boundaries accurately, and the boundaries are defined by K (the siren). The reference turns may be initialized from one such pull/pull of each word spoken in isolation.The total parity may then be set to a convenient constant at the reference seven turns. The training materials are then
It can include utterances with word and segment boundaries as found by the recognition process that represent the utterances to be recognized.

After the statistical data containing a suitable number of training utterances have been accumulated, the reference turns so found are used in place of the initial reference pattern and then the second turn according to the training material. The process will be carried out. The words are then divided into intervals based on decisions made by the recognition processor as in FIG. Each three-frame cover turn (or each reference) (one representative angler cover turn for a turn) is associated with a reference pattern by the pattern matching method described above.

The average values and parances are accumulated for one second so that they form a final set of reference patterns that are derived in a manner fully compatible with the method used by the recognizer.

Preferably, during each training bus, training crazes that are not correctly recognized by the recognition processor are ignored. This is because erroneously recognized utterances may have set the interval boundaries incompletely. The first phrase that was misrecognized upon completion of that training pass was
A new reference pattern can be tried again, and if recognition is successful then, the reference pattern can be further updated.

An alternative to ignoring misrecognized phrases is to create multiple food templates for each training utterance. This template is a combination of templates for each word being uttered in the correct order. The speaker is prompted by the script to speak the instructed word sequence, and the recognition processor only refers to the multiple templates and silent templates. The classification of word boundaries and reference patterns will then be optimal for the given script and available reference patterns. A disadvantage of this procedure is that it may require multiple tests with training scripts.

In order to obtain the highest possible m-translation accuracy, it is preferable to start the training procedure with a set of speaker-independent reference patterns previously determined for the words to be recognized.

Speaker-independent patterns are obtained from phrases representing the phrase to be recognized spoken by at least several different speakers. Word boundaries may be determined by human examination of recorded speech waveforms. Naruko's two-step procedure is then employed to generate speaker-independent patterns. That is, in the first pass, subintervals are uniformly spaced within each word. In the second pass, sub-intervals are determined by a recognition process using the reference pattern from the first bus. Global statistics for all speakers are derived on each pass.

The system is advantageously trained for a particular speaker using the previously generated speaker-independent patterns and, in combination with silent templates, delimits the speaker-dependent speech input. It is possible to make a decision. Preferably, the speaker-dependent speech input is provided in continuous rather than isolated form. More accurate results can be obtained by using continuous audio in the training process. In this way, the boundaries of speaker-dependent speech are determined using the speaker-independent reference patterns available to the device, and the multi-testing process described above for training the device is used;
That is, during the first pass uniformly spaced sub-intervals are established in each word, and in the second pass the recognition process determines the sub-intervals using the pattern generated by the first pass. be done.

Surprisingly, K can conveniently adopt a similar method for previously unknown spoken words.

That is, the boundaries of unknown word words are defined by (1) patterns of speaker independence relative to other word words for recognizing unknown keywords and - (2) occurrences of silence at the beginning and end of the word. Determined using squiggle knowledge to define the limits of the word. The boundaries are then determined by a relatively good score formed to match the speaker-independent reference pattern to the unknown word word rather than "silently" matching it. This result can be used to set boundaries for unknown candidate words, and then the two-step method described above can be employed. That is, during the first pass, the word is evenly divided into subintervals to obtain global statistics, and then during the second pass, the standard recognition process and the reference pattern generated during the first pass are used. That's what I do. This automatic machine method works better than setting unknown words, for example, by a human.

What we wish to make clear is that a "silent" recognition method that uses at least two silent spellings, one of which is preferably dynamically determined, has significant implications in the context of training the device on new speakers. It is about bringing benefits. In this connection, we would also like to point out that the silent "word" acts as a control word for triggering a response from the device. Other control words could be employed if their recognition is sufficiently ICm-realistic; and in some situations, multiple control words could be used to act as "signposts" during a recognition process. However, in the preferred embodiment, silent "words" are the only control words used.

The minimum (required) and maximum (required and optional) dwell times are preferably determined during the training process. In a preferred embodiment of the invention, the device is trained using several speakers as described above. Furthermore, as mentioned above, in the present recognition method, during the training procedure, the boundaries of the pattern are automatically determined according to the method described above. In this way, the boundaries are recorded and the dwell time is stored for each keyword classified by the device.

At the end of the training process, the dwell times for each pattern are tested and the minimum and maximum dwell times for the pattern are selected. In a preferred embodiment of the invention, a histogram of dwell times is formed, and the minimum and maximum dwell times are , No. 25 and No. 75.1
Set to 00 quantile. This gives high recognition accuracy while maintaining a low false alarm rate. Other choices of minimum and maximum dwell times are possible instead, but there is a trade-off between recognition accuracy and false alarm rate. That is, if a minimum dwell time and a maximum dwell time are selected, higher recognition accuracy is generally obtained at the expense of a higher false alarm rate.

Next, the syntax processor will be explained.

The combination of two or three specific word templates is a common example of syntactic control in the decision process.

Referring to Figure 9, odd numbers (1°3.5, 7...)
The syntax circuit 308 for detecting word sequences containing words has two independent sets of pattern alignment registers 310, 312 maintained for each word. The score that falls into the first template is the score for silent or the best score of the set of second templates, whichever is better. The score that falls into the second template is the best score of the first set of ρ templates. This score is also sent to a second silent detection template at node 313Vc. node 313
When a terminal silence of an utterance is detected, as measured by the door template, the label and duration of the uttered word are alternately traced in the first and second set of templates. What is important is that the position of the silent detection template ensures that only silents after a word string with an odd number of words can be detected.

A slightly more complex syntax network is node 313 in the 4$9 diagram.
This can be done by associating each syntax node such as a with a list of acceptable word string lengths. For example, in the 9E syntax network that accepts any string containing an odd number of words, the string length can be set to a particular odd number, e.g., 5, by manually testing the string length in the second silent register 313a. Can be fixed. If the string length at that point is not 5, the register will be inactive (for that analysis interval) and no string score will be reported from that register, but if the string length is 5 then the detection of the string will be reported. can be done. Similarly, the first word 11g register 310 may be enabled when the incoming string length is 0.2 or 4, and the second register may only be enabled when the incoming string length is 1 or 3. To obtain optimal results for a 5-word string, a total of 5 sets of dynamic programming accumulators would be required, but our method allows for a smaller number of accumulators with a slight improvement in recognition accuracy. Multiple roles can be performed simply by providing a reduction.

In the particular embodiment presented, this FfAaltKN includes:
It is designed to recognize either 5-digit strings or known word words that are not digits. This textual syntax is illustrated in Figure 9A. 991, each node (node) 314a s 3 '14 b
% -- 314 h represents a stage in the 1 edict process. Nodes 314a and 314g represent silent recognition, and nodes 314 b z 314
c s 314 d % 3.114 e and 314
f represents recognition of numbers, and 314h represents recognition of non-silent non-number word words. Thus,
According to the syntax control of the device, node 314a
The corresponding silent must first be recognized. At this point, the control is /-)'31 from the number 'N11lk.
4 Moves to bK, control is controlled by node 3 due to non-numeric recognition
14hK Transition (where "transition" refers to an acceptable or "legal" progression in grammatical syntax)
.

At node 314b, the only acceptable progression away from this node is to node 314c, which is a numeric node. On the other hand, at 314h, the only acceptable progression away from this node is to node 314g, which is silent. These include the control syntax processor 308 described in connection with FIG.
is the only acceptable or lawful proceeding allowed by. Importantly, the syntax processor in FIG. 9A, as in FIG. A considerable simplification can be achieved by controlling the progression through a "folded" or "folded" syntax clause structure. The MeA diagram can thus be reconstructed as shown in FIG. 9, provided that certain limitations are placed on the transition from one node to another along the tangent M line section.

A collapsed syntax clause structure is schematically illustrated in FIG. 9B. In this figure, the node 314χ is the only silent node, and the nodes 314u, 31
4v and 314W are 1 new number nodes (old node 3141), 314cs 314ds 314e,
and 314 fK compatible), and node 314h is
A non-numeric node and not a silent node. Silent Note plays a "double role" here. That is,
The silent node 314x represents either silence at the beginning of word string recognition U6 or silence at the end of word string recognition.

Similarly, nodes 314u and 31.14'/ also play dual roles, with one node 314u representing either the first or fourth digit of the word string, and node 314u representing either the first or fourth digit of the word string.
v represents a second or third digit.

In operation, input to each node is accepted according to the digitword count.

The nodes in FIG. 9B represent calculations proceeding in parallel for alternating assumptions. The arc lines represent the mutual dependence of alternating assumptions. In Figure 9B, the 9th AgJ
Assume ih3 instead of the five numbers assumed in
Only one number remains active. In operation,
The assumed activity digit reduction is only if the input arch has a valid word count in relation to the data, i.e. an acceptable word count of 1 from a set of alternative word counts for that arc. This is achieved by accepting human arc data only when the data is available. Hidden 11 node 3
14u accepts input arc data from node 314x only when the word count value associated with the data is 00;
This will always be the case since the data on all arcs exiting the silent node will have the word count set to zero. Node 314u also accepts input arc data from node 314W when the word count value associated with the data is 3.O node accepts the best scoring data from all acceptable inputs. Thus, node 314u can either assume that the digit is becoming consistent with the first digit being uttered or that the digit is being uttered, depending solely on whether data is selected from node 314x or from node 314W. It represents one of the assumptions that it matches the fourth number in the table. Similarly, a silent node accepts arc data from node 314v when node 314v has an associated word count of five. The silent node also accepts input from node 314h and from itself, i.e., node 314.The silent node then chooses the best score data from these acceptable inputs. An advantage of providing structure is to reduce memory requirements and computational load on the G11 device. On the other hand, by discarding certain data and forcing decisions, there is a risk that bad information will be discarded and incorrect decisions will be made. However, if the recognition accuracy is high as in the device described below, the possibility of discarding good data is very low.

If we discard the input from node 314X in favor of the input from 314W, then the very low probability data from the silent note K will be discarded. This is the preferred method of operation, since at any given time the device only needs to determine whether the string is just beginning or whether it has already spoken three words.

The probability of making an error in this judgment is extremely low. The wrapping or folding syntax system requires one additional register per node to maintain the total number of words recognized. (In a slightly more general case, the count value could be the number of words that are &[ in the grammatical syntax string 0)L However, the benefits of the &[ syntactic system, i.e. the benefits of reduced memory and computation. outweighs the drawbacks mentioned above. Another advantage of using syntax in keyword recognition is that the decision whether silence occurs or not is made using squiggle knowledge (grammatical syntax). This syntax allows the device to be ``silent.''
can be detected more reliably, and continuous word strings and “
The boundaries between "silent" can be precisely defined. An important element of the method of the invention is the detection of silences in combination with word strings. That is, when the silent "spelling" K correspondence score corresponds to the recognition of a word string that matches the requirements of grammatical syntax, it contains the "good likelihood score" of the previously received speech signal, so the word Silence is reliably detected at the end of the string.

It is the silent determination of its syntax that allows more accurate and reliable recognition to be made. This is a clear advantage over, for example, methods that recognize silence as an amplitude minimum regardless of audio syntax.

Next, a device implemented using this speech recognition method will be described.

In a preferred embodiment of the invention, signal and data manipulations beyond those performed by the preprocessor of FIG.
228,498 in combination with a special purpose vector computer processor such as that described in US Pat. No. 228,498.

The method of the present invention can be implemented using hardware in addition to using computer programming.

In operation, the apparatus 10 of the present invention operates according to dynamic programming techniques.

The likelihood score sequence for each random likelihood score sequence, ie, the likelihood score sequence for each reference pattern sequence in a known predetermined order, is transmitted from the computer via line 320 to one of memories 322 and 324.
one existing score.

The memory includes (a) a syntax processor 308 that receives a score corresponding to the end of each possible word; (b) a minimum score that supersedes the output of memories 322 and 324 depending on the memory selection and the next phoneme signal. register 3
26, and (C) function alternately as follows under the control of other controls and clock signals.

In operation, the circuit operates according to rules for updating the registers corresponding to each day in FIG. 4 and providing a decision mechanism that can achieve the best match at each pause or silent recognition.

Memories 322 and 324 have the same form;
It is exchanged every 10 milliseconds, ie each time a new frame is analyzed. The memories each have a plurality of 32-bit words, one and the number of 32-bit words being stored in the register associated with one machine word (i.e. the fourth
The circle in the figure corresponds to -1. Initially, one memory, e.g. 322, stores a "bad" likelihood score, ie a score with a large value in this example. Thereafter, the memory 322 is sequentially read out in a predetermined order corresponding to the new likelihood score ordering provided by the vector vector processor via il [320] and the scores are updated as described below. The other memory 3241fC is rewritten. In the next 10 millisecond frame, the now stale score is read from memory 324 and written to another memory 322. This alternating function continues under the control of the syntax processor, minimum score register 326 and other control and clock signals. As mentioned above, memory 322 and 32
Each word of 4 is a 32 bit number. The lower 16 bits, bits 0-15, are employed to store the cumulative likelihood score. Also, bits 16-23 are used to record the phoneme duration, and bits 24-31 are used to record the phoneme duration.
It is employed to store the word duration in that register.

The likelihood score coming from the computer is stored in turn score memory 328 for each frame time.

This information is provided in bursts and sequences from the computer at very high data transfer rates, and then read from the pattern score memory at the low speed employed in the circuit of FIG. In the absence of intervening control from the minimum score register, the output of the selected memory 322 or 324 is fed through the corresponding selected gate 330 or 332 to line 334. a334 is the likelihood score, phoneme or Adder 336.338.340 for updating target pattern duration it value and word duration count value respectively
It is connected to the. Thus, the likelihood score corresponding to the previous frame's score coming from one of the memories 322, 324 is output from the pattern score memory via line 342, added to the old likelihood score, and used for writing. Not stored in memory. The memory selection function is provided by the signal level on line 344. At the same time, the word and phoneme duration counts are incremented by one.

Similarly, word duration counters, phoneme duration counts, and likelihood scores are typically updated.

Two exceptions to the normal update rule described above occur in response to the start of a new phoneme and the start of a new word. At the start of a new speech (which is not at the start of a new word) the first register of a phoneme is not updated according to the usual rules, but instead the likelihood score on @342 is The dwell time register or the preceding phoneme is added to the minimum score from the last register of required dwell times. This is implemented by employing a minimum score register 326. The output of the minimum scorer register represents the minimum score in the previous frame time rIIIVC for the previous phoneme. This score is obtained by continuously updating the contents of the minimum score register as new scores are provided. New l
/1 min score & is loaded into the min score register by taking the sign bit output of the subtraction operation element 346. Element 346 compares the current minimum score with the new minimum score from the updated register.
The minimum score register is further output on line 334 at the beginning of a new phoneme along with the word duration count value corresponding to the register with the minimum score. This output process is enabled at the beginning of a new phoneme. The gates are controlled using a combination of control signals that universalize gates 332 and 330 during the initiation of a new phoneme.

Syntax processor 308 (corresponding to Figure 9B)
is employed to update the first register of the first phoneme for the new word with the best score of the word taking into account the syntax of the word ending in the previous frame.

Thus, when the score of the register corresponding to the first register of a word (new) is updated by the incoming likelihood score, it is not the output of one of the memories 322, 324 that is adopted instead. The best likelihood score of the word ending in the previous frame, preferably taking into account the syntax, is utilized. This function is performed by gate 330
and 332 and simultaneously enable gate 350 to provide the best available score stored in register 352 on line 334 and add it with the incoming turn likelihood score on line 342. becomes.

In this manner, each register corresponding to the dwell time of the reference frame is continuously updated in this hardware implementation. Once a silent word is represented by a likelihood score, the syntax processor provides the necessary control system to enable hardware or computer equipment to perform a silent trace to determine the recognized word. designed to.

Upon consideration of the foregoing description, it will be seen that the various objects of the present invention have been achieved and the beneficial effects have been achieved.

It will be appreciated that the word string continuous speech recognition method and apparatus disclosed herein includes recognition of isolated speech as a particular application. Additions, deletions, and modifications to the embodiments disclosed herein will be apparent to those skilled in the art within the scope of the following claims.

[Brief explanation of the drawing]

FIG. 1 is a flowchart illustrating in general terms the sequence of operations performed in accordance with the method of the invention; FIG. 1A is an electrical block diagram of the apparatus of a preferred embodiment of the invention; FIG. FIG. 3 is a flowchart of a digital computer program for carrying out specific processing operations in the process of FIG. 1; FIG. A diagram of the pattern alignment process of the invention ＼FIG. 5 shows the likelihood function processor of the preferred embodiment of the invention! 6 is a schematic electrical block diagram of a subtraction/absolute value circuit according to a preferred embodiment of the present invention, and FIG. 7 is an electrical block diagram of an overflow detection circuit according to a preferred embodiment of the present invention.
Electrical circuit of logic circuit - Figure 8 is the true value of the circuit diagram in Figure 7 ≦
, FIG. 9 is a schematic flow diagram of a syntax processor of one preferred embodiment of the preprocessor of the present invention, FIG. 9A.
Figure 9B is a schematic flow diagram of a syntax processor that recognizes five-digit word strings delimited by silents; Figure 9B is a schematic flow diagram that folds the flow diagram of Figure 9 to reduce the number of nodes. 10 is an electrical circuit diagram of a sequential decoding pattern and alignment circuit of a preferred specific embodiment of the present invention. 13: Mu/D converter 45: Control processor 46: Preprocessor 48a: Vector processor 48b: Likelihood function processor 49: Sequential decoding processor 51: Clock oscillator 52: Frequency divider 53: Latch 56: Digital multiplier 58: 32 Word circulation Shift register 59: Multiplexer 60: B selection circuit 63: 32 Word shift register memory 65: 32
Bit adder 67: Gate 71: Computer interrupt circuit 73: Inter 7 Ace agent's name: Motoi Kurauchi, Hirodo Kurahashi, Ei drawing no puri? )(
No change in content (method) May 24, 1980 Commissioner of the Japan Patent Office Kazuo Wakasugi Display of the case Patent Application No. 552 of 1980 Title of the invention Amendment for improvement of continuous speech recognition Relationship with Patent No. 11 “Ban-Tsuna-Daki Ami-shou H4 Shoho Tsuzuke Ryuji Rin→- Power of attorney and its translation 1 copy each Drawing
1 Contents of amendments to the detailed description of the invention and brief description of the drawings in the specification Reprint of the drawings as attached (no changes to the contents) The detailed description of the invention and the brief description of the drawings in the specification are as follows: I will correct it accordingly. 1. On page 27 of the specification, lines 10-11 and lines 14-15, "Figure 10" is corrected to "Figure 9." 2. On page 65, line 4, “Figure 8” was replaced with “
0.5 On page 65, line 5, insert the table below after the words ``represents.''"1 1 1 0 1 1 0 0 1 0 1 1 (overflow) 1
0 0 0 0 1 1 0 o 1o 1 (over 70-) 0
0 1 0 0 0 0 0” 4. On page 83, line 15, “Figure 9” is corrected to “Figure 8.” 5 "Figure 9" on page 84, lines 15 and 19
The text has been corrected to read "Figure 8." -6, on page 86, lines 1 and 2, “Figure 9A” was replaced with “Figure 8
I corrected it to ``Figure A''. Z On page 87, line 1 of the same page, "Figure 10" was replaced with "
Figure 9” is corrected. 8. On page 87, lines 4 and 10, "Figure 9A" is corrected to "Figure 8A." 9 Figures 1-9 on page 87, lines 5 and 12.
The text has been corrected to read "Figure 8." 10. On page 87, line 14, "Figure 9B" is corrected to "Figure 8B." 11. On page 88, line 14 of the same page, “Figure 9B” has been corrected to “Figure 8B 1.” 12. On page 88, line 15 of the same page, “Figure 9A” has been changed to “8th AWJ.” I am making a correction. 14 On page 97, line 14 of the same, "Figure 9B" will be corrected to "Figure 8B." 14. On page 100, line 6 of the same page, ``In Figure 7...No. 9
Delete the text that says "The figure is." 15. On page 100, lines 5 and 9, "Figure 9A" is corrected to "Figure 8A." 16. On page 100, line 8 of the same, "Figure 9B" is corrected to "Figure 8B." 1Z On page 100, line 10, "Figure 10" is corrected to "Figure 9."

Claims (1)

Claims: (1) A method of recognizing silence in an incoming voice signal in a voice analysis device that recognizes at least one keyword in the voice signal, comprising: generate a second target template to convert the incoming audio signal into the first target template;
and a second target template, generating a numerical value representing the result of the comparison, and determining whether silence has been detected based at least on this numerical value. (2) The method of claim 1, wherein the step of generating a target template includes adding a dynamically changing silent target template to the incoming audio signal with respect to the first and second target templates. Silent recognition methods that involve occurring in response. (3) In a method for recognizing silence in an audio signal in a speech analysis device that recognizes at least one keyword in an audio signal, the possibility (likelihood) that a currently incoming audio signal portion corresponds to a reference pattern representing silence.
, effectively modify this value according to a syntax-dependent measure representing the recognition according to the grammatical syntax of the immediately preceding part of the audio signal, and from the validly modified value, calculate the current signal. A silent recognition method characterized by determining whether a part corresponds to silent. (4) a reference pattern representing at least one keyword in a speech analysis device that recognizes at least one keyword in a sequence of audio signals, each characterized by a template having at least one target pattern and tailored to the speaker; forming a speaker-independent reference pattern representing the keyword, and using this speaker-independent reference pattern to determine the boundaries of the keyword in the audio signal spoken by the speaker. (5) Patent 5. The method of claim 4, wherein: 1 the training step divides an incoming speech signal representing keywords from a speaker into a plurality of sub-intervals using the keyword boundaries; forcing patterns to correspond, repeating the segmentation and matching steps for multiple audio input signals representing the same keyword, generating statistical data describing the reference pattern associated with each sub-interval; A method of silent recognition comprising making a second pass through the audio input signal representing the keyword using data to generate by a device a sub-interval for the keyword. (6) A method of forming a reference pattern representing a keyword unknown in advance in a speech analysis device that recognizes at least one keyword in an audio signal each characterized by a template having at least one target pattern, the method comprising: forming a speaker-independent reference pattern representing a keyword of A method for forming a reference pattern, comprising training a speech analysis device using the determined boundaries to generate statistical data describing the previously unknown keyword. (7) A method as claimed in claim 6, in which a reference pattern is provided for providing in isolated form an audio signal representing the unknown keyword spoken by the speaker. (8) The method according to claim 6, wherein the training step divides the incoming audio signal corresponding to the previously unknown keyword into a plurality of sub-intervals using the boundary, and each sub-interval, - forcing the pulses to correspond to a unique reference pattern, and repeating this step of splitting and forcing correspondence to multiple audio input signals representing the same keyword (repeated at A, the reference associated with each sub-interval) for each turn. The basic method is to generate statistical data to describe the keyword, use the generated statistical data to generate a second node based on an audio input signal representing the previously unknown keyword, and generate a subroutine for the keyword by the device. Pattern formation method. (9) Each keyword has at least one target keyword (
speech that recognizes multiple keywords in an audio signal, characterized by a template with turns, and described by a grammatical syntax in which each series of keywords in the audio signal is characterized by a plurality of connected series of decision nodes; In the speech analysis method in the analysis device, dynamic programming is adopted to generate a series of numerical scores for recognizing keywords in the speech signal, and the grammatical syntax is adopted to determine which score is used in the recognition process. (10) A speech recognition method characterized in that the number of decision nodes is reduced by determining whether an acceptable progression is formed and collapsing the syntax, thereby reducing the computational load on the device. (10) Speech In an apparatus for recognizing silence in an incoming voice signal in a voice analysis apparatus for recognizing at least one keyword in the signal, means for generating at least first and second target templates representing alternating silent instructions in the incoming voice signal. and the incoming audio signal to the first and Jl! 2 - means for generating a numerical value representative of the result of this comparison; and means for determining, based at least on said numerical value, whether a silence is detected and is a daughter. Silent recognition device. (a) The device according to claim 10,
y) R recognition apparatus, wherein said generating means includes means for generating a dynamically changing target template in response to said incoming human voice signal for one of said first and second target templates. (2) At least one key word in a speech signal is detected in a speech analysis device that recognizes silence in a speech signal.In a device that recognizes silence in a speech signal, there is a possibility that the currently incoming speech signal portion corresponds to a reference snoring representing silence. means for generating a numerical value (degree); means for adding to said numerical value a syntax-dependent value representing recognition according to the grammatical syntax of the immediately preceding portion of said audio signal to form a score; and means for determining whether a portion corresponds to silence. a3 In a speech analysis device for recognizing at least one keyword in a speech signal, each characterized by a template having at least one target pattern, in a device for forming a reference pattern representative of said keyword and tailored to the speaker; , means for forming a speaker-independent reference pattern representing the keyword; and using the speaker-independent reference pattern to determine boundaries of the keyword in an audio signal spoken by the speaker. and means for training a speech analysis device on a speaker using boundaries determined by the device for the keywords spoken by the speaker. Pattern forming device. α◆ The apparatus of claim 13, wherein the training means iteratively divides an incoming speech signal representing a keyword from the speaker into a plurality of sub-intervals using the keyword boundaries. means for iteratively forcing each sub-interval to correspond to a unique reference pattern; means for generating statistical data describing the reference pattern associated with each sub-interval; and means for generating statistical data describing the reference pattern associated with each sub-interval; a reference pattern forming apparatus, comprising means for performing a second pass with the audio signal representative of the keyword using the apparatus to generate a subinterval for the keyword; (v) A device for forming a reference pattern representing a keyword unknown in advance in a speech analysis device that recognizes at least one keyword in an audio signal each characterized by an on-plate having at least one target pattern, the device forming a reference pattern representing a keyword unknown in advance to the device; means for forming a speaker-independent reference pattern representing a keyword; means for determining boundaries of said unknown keyword using said speaker-independent reference pattern; and means for training a speech analysis device using boundaries predetermined by said device to generate statistical data describing said unknown keyword. aQ In the device according to claim 15,
A reference patterning device comprising means for providing in a discrete manner an audio signal representative of the unknown keyword spoken by the speaker. αυ The device according to claim 15! In,
The training means includes means for iteratively dividing the incoming audio signal corresponding to the previously unknown keyword into a plurality of sub-intervals using the boundary; and recursively converting each sub-interval into a unique reference pattern. means for generating statistical data that is described based on the reference pattern associated with each sample interval, and using the generated statistical data to generate the previously unknown keyword. a reference pattern forming device comprising means for making a second pass with said audio signal representing said keyword and generating by said device a sub-interval for said keyword; Number Each keyword must be at least 7. - a plurality of keywords in the speech signal characterized by a template with one target pattern, each series of keywords in the speech signal being characterized by a plurality of connected decision nodes; means employing dynamic programming to generate a set of numerical scores for recognizing keywords in said audio signal; and employing said grammatical syntax to determine which scores are acceptable progress in the recognition process. A speech recognition device comprising: means for determining whether to determine K; and means for reducing the number of decision nodes and reducing the computational load on the device. QcJ A plurality of keywords in an audio signal described in a grammatical syntax, where each keyword is characterized by a template with at least one target pattern, and a sequence of keywords in the audio signal is characterized by a plurality of connected decision nodes. employing the grammatical syntax and means for providing a series of numerical scores for recognizing keywords in the audio signal using circular dynamic programming in a recognition device in a speech analysis device that recognizes keywords; means for determining which scores constitute an acceptable progression in the recognition process; and means for suspending acceptable progression using increments, thereby causing normally acceptable progression to be abandoned in accordance with said syntax. A speech recognition device comprising: (Translation) An audio signal in which each keyword is characterized by a template having at least one target pattern, and a sequence of keywords in the audio signal is described by a grammatical syntax characterized by a plurality of connected decision nodes. A speech recognition method in a speech analysis device that recognizes a plurality of keywords employs dynamic programming to generate a series of numerical scores for recognizing keyword representations in the speech signal, and employs the grammatical syntax. to determine which scores form an acceptable progression in the recognition process, and use the increment to suspend acceptable progression and to abandon normally acceptable progression according to the syntax above. Characteristic voice recognition method.