JPS6228479B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a speech recognition method, and more particularly to a method for recognizing one or more keywords in a continuous audio signal in real time. Various speech recognition systems that recognize isolated utterances by comparing a suitably processed unknown isolated speech signal with a pre-prepared representation of one or more known keywords. has been proposed so far. In this context, the term "keyword" is used to mean a group of connected phonemes and sounds, such as syllables, words, phrases,
It is a part of etc. Although many systems have met with only moderate success, one system has been successfully used to recognize isolated keywords in commercial applications. This system operates substantially according to the method described in U.S. Pat. No. 4,038,503, and if the boundaries of unknown audio signal data are silence or background noise as measured by this recognition system. For example, it provides a successful method for recognizing one of the keywords in a limited vocabulary.
This system relies on the assumption that the period during which the unknown speech signal occurs is well defined and contains a single utterance. In a continuous speech signal, such as continuous conversational speech, where the keyword boundaries are not known or marked a priori (the isolated words are an aspect of the continuous speech signal), the incoming speech Several methods have been proposed to partition the data, i.e. to determine the boundaries of linguistic units such as phonemes, syllables, words, sentences, etc., before the start of the keyword recognition process. ing. However, these conventional continuous speech systems have had only limited success because no satisfactory segmentation process has been found. Moreover, there are substantive problems. For example, only limited vocabularies can be consistently recognized with low false alarm rates, that recognition accuracy is very sensitive to differences between the speech characteristics of different speakers, and that these systems It is very sensitive to distortions in the audio signal being analyzed, such as typically occurs in audio signals transmitted through telephone communication equipment. Therefore, continuous speech can be easily identified by an observer and
Even if they can be understood, machine recognition of even limited vocabulary keywords in continuous speech signals has not yet achieved great success. A speech analysis system that is effective in recognizing keywords in continuous speech is described in the applicant's US patent application entitled ``Continuous Speech Recognition Method''. The system is characterized by a template, where each keyword consists of a directed sequence of one or more target patterns, and where each target pattern consists of a plurality of temporally separated short-term keywords. Use a method to represent the power spectrum. At the same time, the target pattern covers all important acoustic events in the keyword. The invention described in the above-referenced U.S. patent application repeatedly measures a set of parameters that determine the short-term power spectrum of an audio signal in each of a plurality of equal duration sampling periods, thereby providing successive time commands. generating a sequence of short-term audio power spectral frames; and repeatedly selecting one first frame and at least one subsequent frame from the sequence of short-term power spectral frames to form a multi-frame spectral pattern. It is characterized by a frequency analysis method consisting of steps. The method further includes the step of comparing each multi-frame pattern formed as described above to each first daily pattern of each keyword template, preferably using a likelihood statistic; and determining whether the keyword template corresponds to one of the first target patterns. According to the decision step, for each multi-frame pattern corresponding to a first target pattern of possible candidate keywords, the method selects later-occurring frames to form the later-occurring multi-frame pattern. Features. The method then includes, in a similar manner, determining whether each subsequent multi-frame pattern corresponds to a subsequent target pattern of possible candidate keywords, as well as the multi-frame pattern of the selected sequence. identifying candidate keywords when each corresponds to a target pattern of a keyword template, referred to as a selected keyword template. Although the method described in the above-mentioned US patent application is significantly more effective than prior art systems in recognizing keywords in continuous speech, this method still fails to achieve the desired objective. Therefore, the main object of the present invention is a speech recognition method with improved effectiveness in recognizing keywords in continuous, unmarked speech signals. Another object of the invention is a method adapted to improve the effectiveness of a given system, especially the system's discrimination against false alarms. Another object of the invention is a method that responds sufficiently equally to different speakers and therefore to different speech characteristics, is reliable, operates in real time, and is compatible with existing recognition methods. It is. Another object of the present invention is a method which is relatively insensitive to phase and amplitude distortions of unknown audio input signal data, and which method is relatively insensitive to phase and amplitude distortion of unknown audio input signal data.
It is a method that is relatively insensitive to changes in velocity and the dimensionality of the unknown input signal.
This is a method of reducing The present invention relates to a speech analysis system for recognizing at least one predetermined keyword in an audio input signal. Each keyword is characterized by a template consisting of one or more target patterns of commanded sequences. Each target pattern represents at least one short-term keyword power spectrum, or frame. At the same time, the target pattern covers all important acoustic events in the keyword. The present invention includes the steps of selecting a sequence of patterns, each of which consists of one or more frames, and the selected sequence of patterns corresponding respectively to a sequence of target patterns of a keyword template, referred to as a selected keyword template. The method features a frequency analysis method that includes identifying candidate keywords when performing a search, and applying a post-decision processing method to improve false alarm rates. The post-decision processing method, in one aspect, includes the steps of normalizing the time intervals between the selected patterns corresponding to the candidate keyword target pattern, and the normalized time intervals for the candidate keywords being imposed by a prosodic test. applying a prosodic test to the normalized time interval, which must meet a timing criterion determined by the normalized time interval. If this test is not satisfied, the candidate keyword is not accepted as a recognized keyword in the example embodiment. In a preferred embodiment, the timing criterion consists of applying a likelihood statistic function to a normalized interval and accepting a candidate word if this likelihood statistic exceeds a predetermined minimum value. In a second embodiment, this criterion consists of applying certain predetermined interval limits to each normalized interval and accepting a candidate word only if the normalized interval falls within this certain limit. . Other aspects of the post-decision processing method of the present invention include applying a likelihood statistic function to the sequence of selected patterns corresponding to the candidate word to determine a goodness index for each of the patterns; The method is characterized by accumulating goodness indexes for these patterns and accepting candidate words if the accumulated goodness index exceeds a predetermined minimum value. In one preferred aspect, the present invention relates to a speech analysis system for recognizing at least one predetermined keyword in a continuous, unbounded audio input signal. Each keyword is 1
It is characterized by a template consisting of a directed sequence of one or more target patterns. Each target pattern represents at least one short term keyword power spectrum. At the same time, the target pattern covers all important acoustic events in the keyword. The present invention repeatedly measures a set of parameters that determine the short-term power spectrum of an audio signal within each of a plurality of equal duration sampling periods, thereby determining the short-term power spectrum of the audio signal in a continuous time-commanded sequence. An analysis method comprising the steps of: generating a power spectral frame; and repeatedly selecting one first frame and at least one subsequent frame from a sequence of short term power spectral frames to form a multi-frame spectral pattern. It is a feature. The method further includes the step of comparing each multi-frame pattern formed as described above to each first target pattern of each keyword template, preferably using a likelihood statistic; 1 of the first objective pattern of the template
The method is characterized by a step of determining whether or not it corresponds to the above. According to the decision step, for each multi-frame pattern corresponding to a first target pattern of possible candidate keywords, the method selects later-occurring frames to form the later-occurring multi-frame pattern. Features. The method then includes, in a similar manner, determining whether each subsequent multi-frame pattern corresponds to a subsequent target pattern of possible candidate keywords, and the sequence of selected multi-frame patterns. identifying candidate keywords when each corresponds to a target pattern of a keyword template, referred to as a selected keyword template. The method further features a post-decision processing method step that improves the false alarm rate. In one aspect, the post-decision processing method includes the steps of normalizing the time intervals between multi-frame patterns corresponding to selected candidate keywords, and applying a prosodic test to the normalized time intervals to determine the normalization of the candidate keywords. and ensuring that the standardized time intervals have to meet timing criteria imposed by a prosodic test. If this test is not satisfied, in the illustrated embodiment, the candidate keyword is not accepted as a recognized keyword. In another aspect of the post-decision processing method of the present invention, applying a likelihood ratio test method to the selected sequence of multi-frame patterns corresponding to the candidate word to determine a goodness index for each of these patterns. and accumulating goodness indexes for these patterns; and accepting a candidate word if the accumulated goodness index exceeds a predetermined minimum value. The present invention also repeatedly measures a set of parameters that determine the short-term power spectrum of an audio signal in each of a plurality of equal duration sampling periods, thereby determining the short-term power spectrum of a continuous time-commanded sequence. generating an audio power spectrum frame; repeatedly generating a peak spectrum corresponding to the spectrum frame by a fast attack/slow decay peak detection function; dividing the amplitude of by the corresponding intensity value in the corresponding peak spectrum; selecting a sequence of frame patterns; and selecting a keyword temple in which the frame pattern of the selected sequence is referred to as a selected keyword temple identifying candidate keywords when they respectively correspond to target patterns. Preferably, the invention further comprises selecting the values of each of the peak spectral frequency bands from the maximum value of the incoming new spectral values for these frequency bands and the previous peak spectral value multiplied by a constant decay factor having a value less than one. It is characterized by the stages of Another aspect of the invention features a pattern recognition method for identifying in a data stream at least one target pattern characterized by a vector of recognition elements x _i having a statistical distribution. It is. This aspect consists of a step of determining a covariance matrix K and an expectation vector from a plurality of design setting pattern samples x of the target pattern, and a step of determining a covariance matrix K and an expectation _{vector from a plurality of design setting pattern samples x of the target pattern, and a step of determining a plurality of eigenvectors e i (} where v _i v _i+
1 ) The analysis method is characterized by a step of calculating. The method is further characterized in selecting unknown patterns y from the data stream and converting each pattern y into a new vector W having the form (W ₁ , W ₂ , . . . , W _p , R). Here, W
_i = e _i (y-), p is a positive integer constant less than the number of elements in pattern y, R is the reconstruction error statistic, and

Equal to [expression]. By applying a likelihood statistic function to vector W, it is determined whether pattern y is the same as any one target pattern. In certain aspects of the invention, the method is further characterized by calculating a likelihood statistic according to one or other of the following equations. or L″=-1/2[(R-R) ² /var(R)+l _o va
r(R)] Here, the variable with a bar at the top is the sample mean, and var( ) is the unbiased sample variance. Other objects, features, and advantages of the invention will become apparent from the following description of preferred embodiments, taken in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS Preferred embodiments of the invention will now be described in detail with reference to the accompanying drawings in which corresponding reference characters indicate corresponding parts throughout the drawings. In certain preferred embodiments described herein, speech recognition involves specially configured electronic systems for performing certain analog and digital processing on incoming audio data signals, typically speech, and certain other data signals. The entire apparatus includes both a general purpose digital computer programmed in accordance with the present invention to perform the conversion steps and the numerical measurements. The division of tasks between the hardware and software parts of this system is done in such a way as to obtain an overall system capable of achieving speech recognition in real time at a moderate cost. However, some of the tasks performed in hardware in this particular system could also be performed satisfactorily in software, and some of the tasks performed by software programming in this example are It should be understood that different embodiments of the invention can also be implemented with dedicated circuitry. As mentioned above, one aspect of the present invention is the provision of an apparatus for recognizing keywords in continuous speech signals even though the signals are distorted by, for example, telephone lines. Thus, with particular reference to FIG. 1, the voice input signal over the line designated 10 is generated by a carbon telephone transmitter and transmitted over a telephone line spanning any distance, or over a telephone line connecting multiple exchanges. It can be regarded as a received audio signal.
A typical application of the invention is therefore the recognition of keywords in voice data from unknown sources received over a telephone system. On the other hand, the input signal may be any audio data signal, such as an audio input signal, taken, for example, from a wireless telecommunications link, from a commercial broadcast station, or from a private communications link. As will be apparent from the above description, the present method and apparatus relate to the recognition of speech signals containing sequences of sounds or phonemes or other recognizable markings. The terms "keyword,""sequence of target patterns,""templatepattern," or "keyword template" are used herein and in the claims, but these four terms are used collectively. and is considered to be consent. This is a convenient way of representing recognizable sequences of audio sounds, or representations thereof, that are detectable by the present method and apparatus. These terms should be interpreted broadly and generally to include everything from a single phoneme, syllable, or sound to a sequence of words (in the grammatical sense) as well as a single word. Analog-to-digital (A/D) converter 13 receives incoming analog audio signal data on line 10 and converts the signal amplitude of the incoming data to digital form. The exemplary A/D converter is designed to convert input signal data to a 12-bit binary representation, and this conversion occurs at a rate of 8000 conversions per second. A/D converter 13 provides its output via line 15 to autocorrelator 17. Autocorrelator 17 processes the digital input signal to generate short term autocorrelation functions 100 times per second and provides its output over line 19 as directed. Each autocorrelation function contains 32 values or channels, and each value is computed to a resolution of 30 bits.
The autocorrelator will be described in detail below with reference to FIG. The autocorrelation function through line 19 is Fourier transformed by Fourier transform device 21 and the corresponding short-term wind power spectrum is converted into line 23.
Get on top. The spectra are generated at the same repetition rate as the autocorrelation function, ie, 100 times per second, and each short-term power spectrum has 31 numerical terms, each with a resolution of 16 bits. As will be appreciated, each of the 31 terms in the spectrum represents signal power within a certain frequency band. The Fourier transform device also preferably includes a Hamming or similar window function to reduce spurious adjacent band responses. In the illustrated embodiment, the Fourier transform as well as subsequent processing steps are carried out under the control of a suitably programmed general purpose digital computer using a peripheral array processor to speed up the arithmetic operations repeatedly required in accordance with the method. is executed using The particular combinator used was a model PDP-11 manufactured by Digital Equipment Corporation, USA. The specific sequence processor used is described in US Patent Application No. 841,390. The programming described below with reference to FIG. 3 is substantially based on the capabilities and characteristics of these commercially available digital processing units. The short-term wind power spectrum is frequency response equalized, as indicated at 25, and this equalization is performed as a function of the peak amplitude occurring in each frequency band or channel as described in detail below. Frequency response equalized spectra are generated on line 26 at a rate of 100 per second, each spectrum having 31 numerical terms evaluated to 16 bit accuracy. To facilitate the final evaluation of the incoming audio data, the frequency response equivalent window spectrum in line 26 is subjected to an amplitude transformation, as indicated at 35, which imposes a non-linear amplitude transformation on the incoming spectrum. This transformation will be described in detail later, but it is worth mentioning here that it improves the accuracy with which the unknown incoming speech signal can be matched with the keywords of the reference vocabulary. In the exemplary embodiment, this transformation is performed on the entire frequency response equalization wind spectrum at a time prior to comparing the spectrum with keyword templates representing keywords in the reference vocabulary. The amplitude transformed equalized short term spectrum on line 38 is then compared to the keyword template at 40. The keyword template designated 42 represents the keywords of the reference vocabulary in the spectral patterns to which the transformed equalized spectra can be compared. Candidate keywords are thus selected according to the closeness of the comparison. In an exemplary embodiment, the selection process is designed to minimize the likelihood of missing keywords while rejecting grossly inappropriate pattern sequences. Candidate words (and corresponding accumulated statistics regarding the incoming data) are provided at 46 over line 44 for post-decision processing to reduce the false alarm rate. The final verdict is 4
8 is indicated. Post-decision processing, including the use of prosodic masks and/or sound level likelihood ratio tests, improves discrimination between correct detections and false alarms, as described in detail below. Preprocessor In the apparatus shown in FIG. 2, an autocorrelation function with its inherent averaging is applied to the incoming analog audio data over line 10, generally the digital data generated by an analog-to-digital converter 13 from the audio signal. Performed digitally on the stream. Converter 13 provides a digital input signal over line 15. The digital processing functions as well as the input analog-to-digital conversion are timed under the control of clock oscillator 51. The clock oscillator provides a basic timing signal of 256,000 pulses per second, which is passed through frequency divider 52.
to obtain a second timing signal of 8000 pulses per second. This slower timing signal controls analog-to-digital converter 13 as well as latch register 53. Latch register 53 holds the 12-bit result of the immediately previous conversion until the next conversion is completed. The automatic correlation product is the number contained in register 53 and 32
is generated by digital multiplier 56 which multiplies the output of word shift register 58. Shift register 58 is operated in recirculation mode and is driven by the earlier clock frequency;
As a result, one complete rotation of shift register data is achieved for each analog-to-digital conversion. The input to shift register 58 is taken from register 53 once during each complete cycle. One input to digital multiplier 56 is taken directly from latch register 53, while the other input to multiplier 56 is taken from the current output of the shift register via multiplexer 59 (with one exception, described below). . These multiplications are performed at the higher clock frequency. Thus, each value obtained from the A/D conversion is multiplied by each of the 31 preceding conversion values. As will be understood by those skilled in the art, the signal thereby generated is the time-delayed input signal itself plus 32 different time increments (of which 1
(one has zero delay). To create a zero delay correlation, ie, the power of the signal, multiplexer 59 causes the current value of latch register 53 to be multiplied by itself as each new value is introduced into the shift register. This timing function is designated at 60. As also understood by those skilled in the art, the product from a single transform and its 31 predecessors does not adequately represent the energy distribution or spectrum over a reasonable sampling period. The apparatus of FIG. 2 therefore provides an averaged version of the products of these sets. The accumulation process that performs the averaging is interconnected with adders 65 to form a set of 32 accumulators 32
Provided by word shift register 63.
Thus, each word can be recirculated after being added to the corresponding increment from the digital multiplier. The circular loop passes through a gate 67 controlled by a 1/N divider circuit 69. The 1/N divider circuit 69 is driven by a low frequency clock signal. The divider circuit 69 is accumulated and therefore the shift register 6
Divide the low frequency clock by a factor that determines the number of instantaneous autocorrelation functions that are averaged before 3 is read out. In the illustrated example, 80 samples are accumulated before being read. In other words, N of the 1/N divider circuit 69 is equal to 80. After the 80 transform samples have been correlated and accumulated in this manner, the divider circuit 69 sends the computer interrupt circuit 71 to line 7.
Trigger through 2. At this time, the contents of the shift register 63 are transferred to the appropriate interface circuit 7.
The 32 consecutive words in the registers are presented to the computer via the interface circuit 73 in the ordered order. As will be understood by those skilled in the art, the transfer of this data from the peripheral unit, ie, the autocorrelator preprocessor, to the computer can typically be accomplished by direct memory access procedures. It can be seen that based on averaging 80 samples with an initial sampling rate of 8000 samples per second, 100 averaged autocorrelation functions are provided to the computer per second. While the contents of the shift register are being read to the computer, gate 67 is closed, so each of the words in the shift register is effectively reset to zero, allowing the accumulation process to begin again. The operation of the apparatus shown in FIG. 2 can be expressed in mathematical terms as follows. Assuming that an analog-to-digital converter generates a time series S(t), where t = 0, To, 2To, ..., and To is the sampling period (1/8000 seconds in the example embodiment) ), the exemplary digital correlation circuit of FIG. 2, ignoring start-up ambiguities, can be thought of as computing the autocorrelation function of equation (1): Here j = 0, 1, 2, ...31; t = 80To,
160To, . . . , 80nTo, . . . These autocorrelation functions correspond to the correlation outputs on line 19 in FIG. Referring to FIG. 3, the digital correlator operates continuously to transmit a series of data blocks to the computer at a rate of one complete autocorrelation function every 10 milliseconds. This is indicated by 77. Each block of data represents an autocorrelation function generated from a corresponding partial time. As mentioned above, the exemplary autocorrelation function is provided to the computer at a rate of 100 32 word functions per second. In the exemplary embodiment, processing of the autocorrelation function data is performed by a suitably programmed dedicated digital computer. A flowchart containing the functions provided by the computer program is shown in FIG. However, some of the steps can also be performed by hardware rather than software, and similarly some of the functions performed by the apparatus of FIG. I would like to point out that this can also be done using software. Although the digital correlator of Figure 2 performs time averaging with an autocorrelation function generated on an instantaneous basis, the average autocorrelation function read out to a computer can interfere with the orderly processing and evaluation of samples. It may still contain some unusual discontinuities or non-uniformities. Therefore, each block of data, i.e. each autocorrelation function Ψ(j,
t) is first smoothed in time. This is indicated at 79 in the flowchart of FIG. A preferred smoothing process is one in which the smoothed autocorrelation output Ψs(j,t) is given by the following equation (2). Ψs(j, t)=C ₀ Ψ(j, t), C ₁ Ψ(j, t-T)+C ₂ Ψ(j, t+
T) (2) Here, Ψ(j, t) is the unsmooth input autocorrelation defined by equation (1), Ψs(j, t) is the smoothed autocorrelation output, and j is the delay time, t indicates real time and T indicates the time period (equal to 0.01 seconds in the preferred embodiment) between successively generated autocorrelation functions. Although other values can be selected for the weighting functions C ₀ , C ₁ , and C ₂ , in the exceptional embodiment, 1/
Preferably, it is selected to be 2, 1/4, 1/4. A smoothing function that approximates a Gaussian impulse response with a frequency cutoff of, for example, 20 Hertz can be implemented in computer software, for example. However, testing has shown that the exemplary smoothing function, which is easy to implement, provides satisfactory results. As indicated, the smoothing function is applied separately for each value j of delay. 81, the cosine Fourier transform is applied to each time smoothed autocorrelation function Ψs(j,
t) to generate a 31 point power spectrum. This power spectrum is determined as shown in the following equation (3). Here, S(f,t) is the spectral energy in the band centered at f hertz (Hz) at time t, and W(j)=1/2(1+cos2πj/63)
is the Hamming window function that reduces the sidelobes, Ψs(j,t) is the smooth autocorrelation function at delay j and time t, and f=30+1000(0.0552m+0.438) 〓Hz (4) m= 1, 2, ..., 31 These are frequencies equally spaced by pitches on the "mel" scale. As can be appreciated, this corresponds to an essential pitch (mel scale) frequency-to-axis spacing for frequencies in the bandwidth of typical communication channels of approximately 300-3500 Hz.
As can also be understood, each point or value within each spectrum represents a corresponding band of frequencies. Although this Fourier transform can be performed entirely within normal computer hardware, the process can be significantly faster if an external hardware multiplier or Fast Fourier Transform (FFT) peripheral is used. The construction and operation of such modules are well known in the art and will not be described in detail here. A frequency smoothing function is advantageously incorporated into the hardware fast Fourier transform peripheral in which each of the spectra is frequency smoothed according to the preferred Hamming window weighted function W(j) defined above. This is indicated by block 83 in block 85 corresponding to the hardware Fourier transform implementation. Once a continuous smooth power spectrum is received from fast Fourier transform peripheral 85, a communication channel equalization function is applied to peripheral 85, as described below.
is obtained by determining a (typically different) peak power spectrum for each incoming wind power spectrum from , and changing the output of the fast Fourier transform device accordingly. Each newly generated peak amplitude spectrum y(f, t) corresponding to the incoming wind power spectrum S(f, t)
t) (where f is indexed over multiple frequency bands of the spectrum) is the result of a fast attack, slow decay, peak detection function for each of the spectral channels or bands. The wind power spectrum is normalized with respect to each period of the corresponding peak amplitude spectrum. This is indicated by 87. According to an exemplary embodiment, the value of the "old" peak amplitude spectrum y(f,t-T), determined before receiving the new wind spectrum, is the value of the new arriving spectrum S on a frequency band-frequency band basis.
(f, t). A new peak spectrum y(f,t) is generated according to the following rules.
The power amplitude in each band of the "old" peak amplitude spectrum is multiplied by a fixed fraction, for example 511/512 in the illustrative example. This corresponds to the slow decay part of the peak detection function. Arrival spectrum S
If the power amplitude in the frequency band f of (f, t) is larger than the power amplitude in the corresponding frequency band of the attenuated peak amplitude spectrum, then
The attenuated peak amplitude spectral values for that (those) frequency bands are replaced with the spectral values of the corresponding bands of the incoming wind spectrum.
This corresponds to the sudden attack part of the peak detection function. Mathematically, the peak detection function can be expressed as the following equation (5). y(f,t)=max{y(f,t-T)・(1-E),S(f,t)} (5) where f is indexed over each of the frequency bands and y (f,t) is the resulting peak spectrum and y(f,t-T) is the "old"
That is, it is the previous peak spectrum, and S(f,
t) is the new incoming power spectrum and E is the Decay parameter. After the peak spectrum is generated, the resulting peak amplitude spectrum is
At 89, the newly generated peak spectrum is frequency smoothed by averaging each frequency band peak value of the peak values corresponding to adjacent frequencies. The width of the overall band of frequencies contributing to the average value is approximately equal to the representative frequency separation between the formant frequencies. As understood by those in the speech recognition field, this separation is on the order of 1000Hz. By averaging in this particular way,
Useful information in the spectrum, ie local fluctuations representing formant resonances, is preserved, while the global or global emphasis in the frequency spectrum is suppressed. The resulting smoothed peak amplitude spectrum y(f,t) is equal to the just received power spectrum S
(f, t) as the arriving smoothed spectrum S(f, t)
is used to normalize and frequency equalize the amplitude value of each frequency band of y(f,t) by dividing it by the corresponding frequency band value of the smoothed peak spectrum y(f,t). Mathematically, this corresponds to the following equation (6). Sn (f, t) = S (f, t) / y (f, t) (6) where Sn (f, t) is the peak normalized smoothed power spectrum and f is the peak normalized smooth power spectrum over each of the frequency bands. indexed. This stage is 9
1 is indicated. Thus, a sequence of frequency-equalized normalized short-term power spectra results that accentuates changes in the frequency content of the incoming speech signal and suppresses generalized long-term frequency emphasis or distortion. This frequency compensation method is used for the recognition of speech signals transmitted through frequency-distorting communication links, such as telephone lines, in conventional frequency compensation systems where the criterion for compensation is the average power level of the entire signal or each respective frequency band. It was found to be very beneficial compared to It is instructive to point out that although the subsequent spectra have been variously processed and equalized, the data representing the incoming speech signal still consists of spectra occurring at a rate of 100 per second. The normalized, frequency equalized spectrum indicated at 91 is subjected to an amplitude transformation indicated at 93 which performs a non-linear scaling of the spectral amplitude values. Let us denote each equalized and normalized spectrum as Sn(f, t) (from equation (6)), where f indexes different frequency bands of the spectrum and t is the spectrum on a non-linear scale, indicating real time. x(f, t) is a linear fractional function expressed by the following equation (7A). x (f, t) = Sn (f, t) - A/Sn (f, t)
+A (7A) where A is the spectrum defined as
This is the average value of Sn(f, t). Here, f _b indexes over the frequency band of the power spectrum. This scaling function creates a soft threshold and gradual saturation effect on spectral intensities that deviate significantly from the short-term average A. Mathematically, the function is approximately linear for intensities near the average, approximately logarithmic for intensities away from the average, and substantially constant at extreme values of intensity. With respect to the logarithmic scale, the function x
(f,t) is symmetric about zero, and the function exhibits threshold and saturation behavior reminiscent of the auditory nerve stimulation rate function. In fact, the entire recognition system performs much better with this particular non-linear scaling function than with linear or logarithmic scaling of the spectral amplitudes. Thus, the amplitude transformed, frequency response equalized and normalized short-term power spectrum x
A sequence of (f, t) is generated. Here t
teeth,. 01,． 02．． 03．． 04,...equal to seconds and f
=1,...,31 (corresponding to the frequency band of the generated power spectrum). 32 words are provided for each spectrum, and the value of A (Equation 7B), the average value of the spectral values, is stored in 32 words. The amplitude transformed short-term power spectrum is, in the illustrated example, as indicated at 95.
Stored in a first-in-first-out (FIFO) circular memory with storage capacity for 256 32-word spectra. Thus, 2.56 seconds of audio input signal is available for analysis. This storage capacity provides the recognition system with the necessary flexibility to select spectra at different real times for analysis and evaluation, thus having the ability to move forward and backward in time as analysis is required. do. Thus, the amplitude-converted power spectrum for the most recent 2.56 seconds is stored in circular memory and available when needed. In operation, in the illustrated embodiment, each amplitude transformed power spectrum is
Remembered for 2.56 seconds. Thus, the spectrum that enters the circular memory at time t ₁ is lost or shifted from memory after 2.56 seconds, since a new amplitude transformed spectrum is stored corresponding to time t ₁ +2.56. The transformed and equalized short-term power spectra passed through the circular memory are compared, preferably in real time, with keywords of a known vocabulary to detect or select these keywords in continuous audio data. Each vocabulary keyword is a template pattern that statistically represents a plurality of processed power spectra formed into a target pattern (referred to as a design setting pattern) of a plurality of non-overlapping multiframes (preferably three spectra). It is expressed as follows. Preferably, these patterns are selected to best represent the important acoustic events of the keyword. The spectra forming the design set patterns were generated for keywords spoken in various contexts using the same system described above to process successive unknown speech inputs on line 10 as shown in Figure 3. be done. Thus, each keyword in the vocabulary has associated with it a generally plurality of sequences of design set patterns P(i) ₁ , P representing one indication of the i-th keyword in some region of the short-term power spectrum. (i) It has ₂ ,... The collection of design patterns for each keyword forms the statistical basis from which target patterns are generated. In an exemplary embodiment of the invention, the design setting pattern P(i)j can be considered as a 96 element array, each consisting of three selected short-term power spectra arranged in serial order. The power spectra forming the pattern should preferably be spaced at least 30 milliseconds apart to avoid spurious correlations due to time domain smoothing. In other embodiments of the invention, other sampling techniques can be implemented to select spectra. However, a preferred approach is to select spectra that are spaced apart by a constant duration, preferably 30 milliseconds, and to space the non-overlapping design setting patterns apart by each period of time that forms the keyword. Therefore, the first design setting pattern _P1 corresponds to a part near the beginning of the keyword, the second pattern _P2 corresponds to a part delayed in time, and so on.
These patterns P ₁ , P ₂ , . . . are serial or sequence target patterns to which the incoming audio data is combined, i.e. keyword templates,
form a statistical standard for purpose pattern
t ₁ , t ₂ , ..., are independent Gaussian variables that allow P(i)j to generate a likelihood statistic between the selected multi-frame pattern and the target pattern, which will be defined later. Assuming that it consists of
Each consists of statistical data. Thus, the target pattern consists of an array whose entries consist of the standard deviation of the mean and the area normalization factor for the corresponding collection of design setting pattern array entries.
More accurate likelihood statistics will be described later. Virtually all keywords have one or more contextual and/or regional pronunciations, and thus one
It will be apparent to those skilled in the art that there is more than one "spell" of design configuration patterns.
Therefore, the above patterned spellings P ₁ , P ₂ ,...
In reality, the keyword with P(i) ₁ , P
(i) ₂ ,...i=1, 2,...,M can be generally expressed. Here, each P(i)j is a possible alternative description of the jth class of design setting patterns, and there are a total of M different spectra for the keyword. Thus, in the most general sense, the target patterns t ₁ , t ₂ , . . . t _i , . . . each represent a plurality of alternative statistical spellings for the ith group or class of design setting patterns. In the exemplary embodiments described herein, the term "target pattern" is used in its most general sense;
Thus each target pattern may have one or more acceptable alternative "statistical spellings." Processing of Stored Spectra The stored spectra at 95 representing incoming continuous speech data are compared with stored templates of target patterns designated at 96 representing keywords of the vocabulary according to the following method. Each subsequent transformed frequency response equalization spectrum is the first spectral portion of the multi-frame pattern;
Here, they are treated as three spectral patterns corresponding to 96 element vectors. In the exemplary embodiment, the second and third spectral portions of the pattern correspond to spectra occurring 30 and 60 milliseconds later (in real time). In the resulting pattern indicated at 97, the first selected spectrum forms the first 32 elements of the vector, the second selected spectrum forms the second 32 elements of the vector, The third selected spectrum forms the third 32 elements of the vector. Preferably, each multi-frame pattern thus formed is transformed according to the following method in order to reduce cross-correlation, reduce dimensionality, and enhance separation between target pattern classes. This is indicated in 99. The transformed pattern in the exemplary embodiment is provided as input to a statistical likelihood calculation, designated 100, that calculates a measure of the likelihood that the transformed pattern matches the target pattern. Pattern Transformation Considering pattern transformation first, using matrix representation, each multi-frame pattern can be represented by a 96×1 column vector x=(x ₁ +x ₂ , . . . , x ₉₆ ). where x ₁ , x ₂ , ..., x ₃₂ are the elements x of the first spectral frame of the pattern
(f, t ₁ ) and x ₃₃ , x ₃₄ , ..., x ₆₄ are the elements x of the second spectral frame of the pattern
(f, t ₂ ), and x ₆₅ , x ₆₆ , . . . , x ₉₆ are the elements x (f, t ₃ ) of the third spectral frame. It has been experimentally observed that most of the elements x _i of vector x have a probability distribution that clusters symmetrically around their mean value, and thus the Gaussian probability density function is designed to correspond to a particular desired pattern. The set pattern closely fits the distribution of each x _i over samples from a particular collection. However, many pairs of elements x _i ,
x _j are highly correlated, so the assumption that the elements of x are mutually independent and uncorrelated is not accepted. Moreover, the correlation between elements arising from different frames in a multi-frame pattern carries information about the direction of motion of the formant resonances in the input speech signal, and this information can be transmitted even if the mean frequency of the formant resonances varies, e.g. from speaker to speaker. too,
remains relatively constant. As is well known, the direction of motion of formant resonance frequencies is an important cue for human speech recognition. As is well known, the effects of cross-correlation among the elements of x can be accounted for by using the multivariate Gaussian log-likelihood statistic. -L=1/2(x-)K ^-1 (x-) ^t +1/2ln‖
K‖ (8A) where is the sample mean of x and K is the matrix of sample covariance between all pairs of elements of x defined by (8B), ‖K‖
represents the determinant of matrix K. K _ij =(x _i − _i )(x _j − _j ) (8B) The covariance matrix K can be decomposed into eigenvector representations by well-known methods. K=EVE ^t (8C) where E is a matrix of K's eigenvectors e _i and V is a diagonal matrix of K's eigenvalues v _i . These quantities are defined by the following relationships: Ke _i ^t = v _i e _i ^t (8D) Multiplication by matrix E represents vector x96
Corresponds to exact rotation in dimensional space. now,
If the transformation vector w is defined as equation (8E), then w=E(x−) ^t (8E) The likelihood statistic can be rewritten as shown in equation (8F) below. Each eigenvalue v _i is the statistical variance of the random vector x measured in the direction of the eigenvector e _i . The parameters K _ij and _i are determined in the illustrated embodiment by averaging the formed multi-frame pattern over multiple observed design setting samples for each of the indicated statistical relationships. Then, it will be decided. This procedure forms statistical estimates of the expected values of K _ij and x _i . However, the number of independent parameters estimated is (96 mean values) + 96 x 97/2 = 4656 covariances. The achievable number of sample observations per statistical parameter is clearly very small, since it is not feasible to collect more than a few hundred design setting pattern samples for one target pattern. The effect of insufficient sample size is that chance fluctuations in parameter estimates are comparable to the estimated parameters. These relatively large variations induce a strong statistical bias on the classification accuracy of the decision processor based on equation (8F), thus allowing the processor to classify samples from its own design-set patterns with high accuracy. However, the measured performance on unknown data samples is very poor. It is well known that reducing the number of estimated statistical parameters reduces the effects of small sampling biases. For this reason, the following method is commonly used to reduce the dimensionality of statistical random vectors. The eigenvectors e _i defined above are ranked so as to form a ranked matrix E ^r of ranked eigenvectors e ^r by reducing the rank of their associated eigenvalues v _i , so that e ^r ₁
are the directions of maximum variance v ^r _i and v ^r _i+1 v ^r ₁ . The vector x- is transformed into a vector w as in equation (8E) (using the ranked matrix E ^r ), but only the first p elements of w are used to represent the pattern vector x. be done.
In this expression, sometimes referred to as "principal component analysis," the effective number of statistical parameters to be estimated is
It's about 96p instead of 4656. To classify the pattern, the likelihood statistic L is computed as in equation (8F), except that it sums from 1 to p instead of from 1 to 96. When principal component analysis is applied to real data, the classification accuracy of the processor increases as p increases until a critical value of p where the accuracy is maximum, after which the poor performance described above is observed at p = 96. It is observed that p decreases as p increases until
(See graphs (a) and (b) in Figure 4). The maximum classification accuracy achieved by principal component analysis methods is still limited by the effects of small sample statistical bias, and the number of components, or dimensions, required is limited by the number actually needed to represent the data. Much more than expected. Furthermore, the performance for design set pattern samples is actually incrementally worse than the performance for unknown samples over a wide range of p. The latter two sources of influence are due to the transformation vector w p
By representing the sample space by components, the remaining 96 p components do not contribute to the likelihood statistics. The regions where the majority of pattern samples are found have been described, but the regions where few samples occur have not been described. The latter region corresponds to the tail of the probability distribution and thus to the region of overlap between different target pattern classes.
Thus, prior art methods exclude exactly the information needed to make the most difficult classification decisions.
Unfortunately, these overlap regions have high dimensionality, thus reversing the above discussion and e.g.
It is infeasible to use a small number of components of w where _i is not the maximum but is the minimum. According to the invention, the unused components w _p+1 ,...
..., w ₉₆ is estimated by the reconstruction statistic R in the following manner. Formula for L (Formula 8F)
The terms that disappear from include the squares of the components w _i , each weighted according to its variance v _i . All these variances can be approximated by a constant parameter c, which can be extracted. Therefore, The sum on the right is exactly the square of the Euclidean norm (length) of vector w′. Here, w′=(w _p+1 ,..., w ₉₆ ) (8H) Vector w ^p is defined by the following equation. w ^p = (w ₁ ,…, w _p ) (8I) In that case, This is because the vectors w, w' and w ^p can be translated to form a right triangle.
The eigenvector matrix E yields an orthogonal transformation, so the length of w is the same as the length of x-. Therefore, it is not necessary to calculate all components of w. The statistics sought to estimate the influence of unused components on the log-likelihood function L are as follows. This is the length of the difference between the observed vector x- and the vector obtained by trying to reconstruct x- as a linear combination of the first p eigenvectors e _i of K. Therefore, R has the characteristics of a reconstruction error statistic. To use R in the likelihood function, R is simply added to the set of transformed vector components and a new random vector (w ₁ , w ₂ , ..., w _p ,
R). Under this assumption, the new likelihood statistic is as follows. Here M=1/2 (R-R) ² /var(R)+1/2lnv
ar(R)(8M) Variables with bars are sample means, var
( ) represents unbiased sample variance. In equation 8L, the value of _i should be zero, and var
(w _i ) should be equal to v _i . However, since eigenvectors cannot be calculated or applied with infinite arithmetic precision, it is best to remeasure the sample mean and variance after transformation to reduce statistical bias in the system caused by arithmetic rounding errors.
This caution also applies to equation (8F). The measured performance of the likelihood statistic L′ in the same maximum likelihood decision processor is shown in graphs (c) and
It is plotted as (d). It can be seen that as p increases, the classification accuracy reaches a maximum value, but this time at a very small number of dimensions p. Moreover, the maximum accuracy achieved is significantly higher than for statistics L that differ only in the absence of reconstruction errors R. As another test of the efficacy of the reconstruction error statistic R,
The same experiment was repeated again, but this time the likelihood relation used was simply: L″=-M (8N) That is, in this case, the area where most of the sample data was present was ignored, and the area where relatively few samples were found was delineated. (e) and (f)) are almost as high as for the statistic L′, and the maximum value occurs in a smaller number of dimensions, p=3. This result shows that the first that any data sample in the space of eigenvectors of p can be accepted as belonging to the target pattern class, and that there is little or no benefit to be gained by making detailed probability estimates in that space. Statistical likelihood calculation The transformed data w _i corresponding to the formed multi-frame pattern x is fed as input to the statistical likelihood calculation.As mentioned above, this processor continuously , and computes a measure of the probability that the unknown input speech represented by the transformed multi-frame pattern matches each of the target patterns in the keyword template of the machine's vocabulary.Typically, each of the target patterns The data have a slightly asymmetric probability density, but nevertheless mean _i and variance var
It is statistically well approximated by a normal distribution with (w _i ). Here, i is the sequential instruction of the element of the k-th target pattern. The simplest implementation of this process assumes that the data associated with different values of i and k are uncorrelated, so that the overall probability density for data x belonging to target pattern k is (logarithmically) do. Since the logarithm is a monotonic function, this statistic can be used to determine whether the probability of matching any one target pattern in a keyword template is higher or lower than the probability of matching a target pattern in another vocabulary, or It is sufficient to determine whether the probability of matching a particular pattern exceeds a predetermined minimum level. Each input multi-frame pattern has its statistical likelihood L(t|k) calculated against all of the target patterns of keyword templates in the vocabulary. The resulting likelihood statistic L(t|k) is interpreted as the relative likelihood of the occurrence of the pattern of interest, called k, at time t. As is well understood by those skilled in the art, ranking of these likelihood statistics constitutes speech recognition insofar as it can be performed from a single objective pattern. These likelihood statistics can be used in different ways in the overall system depending on the final function to be performed. Selection of Candidate Keywords According to a preferred embodiment of the invention, any first
If the likelihood statistic of the multi-frame pattern with respect to the target pattern exceeds a predetermined threshold, the comparison is indicated at 101, 103, and the incoming data is first
In order to determine the local maximum for the likelihood statistic corresponding to the desired pattern of Further investigation will be conducted to determine whether This is indicated in 105. Therefore, the process of iteratively testing the newly formed multispectral frame against all first target patterns is interrupted, and the probability in the sense of statistical likelihood occurring after the first multiframe pattern is interrupted. A search begins for a pattern that most corresponds to the next (second) target pattern for a given candidate keyword. If the second multi-frame pattern corresponding to the second target pattern is not detected within the preset time window, the search sequence ends;
The recognition process begins again at a time immediately after the end of the first multi-frame pattern that identified potential candidate keywords. Thus, after the first multi-frame pattern yields a likelihood score greater than the required threshold, a timing window is provided and the next objective sequentially corresponds to candidate keywords that may have been selected within this time. A pattern must appear that matches the pattern. The timing window may vary depending on, for example, the duration of the phonetic segment of a particular potential candidate keyword. This process continues until (1) a multi-frame pattern is identified in the incoming data for all of the desired patterns of the keyword template, or
(2) Continue until the target pattern cannot be related to any pattern that occurs within the allowed timing window. If the search is terminated by state (2), then the search for a new first spectral frame follows the termination of the first previously identified multi-frame pattern in the next spectrum, as described above. Starting anew. At this level of processing, the goal is to concatenate possible multi-frame patterns corresponding to the target pattern to form candidate words. (this is
107 as directed. ) Therefore, the detection threshold is set loosely so that correct multiframe patterns are rejected, and here, at this acoustic processing level,
The discrimination between correct detection and false alarms is mainly
It is extremely unlikely that the requirement that multiple pattern events have to be detected together would result. Post-Decision Processing Processing at the acoustic level continues in this manner until the incoming audio signal is terminated. however,
Even after keywords have been identified using the likelihood probability test described above, additional post-decision processing tests (indicated at 109) are used to identify incorrect detections while keeping the probability of correct detection as high as possible. It is preferable to reduce the likelihood of selecting keywords that are false (ie, reduce the false alarm rate). For this reason, the output of the acoustic level processor, i.e. the candidate words selected by the concatenation process, is a mask of the prosodic relative timing window and/or the acoustic level processor associated with all target pattern classes. It is further filtered by a likelihood ratio test using information from the setter. Prosodic Masks As mentioned above, during the determination of the likelihood statistics, times of occurrence of multi-frame patterns with local peak values of the likelihood statistics for the active target pattern are found, and in the preferred embodiment candidate keywords. are recorded for each of the selected patterns corresponding to a plurality of consecutive target patterns. These times pt ₁ , pt ₂ ,... for each candidate keyword
..., p _o is analyzed and evaluated according to a predetermined prosodic mask for that keyword to determine whether the time interval between subsequent pattern likelihood peaks meets predetermined criteria. . According to this method, the elapsed time between the peak values of the likelihood statistics, i.e., i=
For 2, 3, . . . , n, p _i -pt _i-1 is first normalized by dividing each elapsed time period by p _o -pt ₁ . The resulting normalized duration is the prosodic mask for the candidate keyword, i.e.
The candidate word is compared to a sequence of normalized period lengths of acceptable ranges, and if the period lengths fall within the selected range, the candidate word is accepted. In an exemplary embodiment, the prosodic mask timing window is determined by measuring elapsed time for sample keywords spoken by as many different speakers as possible. The prosodic patterns are then compared to the statistical sample keyword times using statistical calculations where the mean and standard deviation for each prosodic mask (corresponding to each keyword) is derived from the keyword design setting pattern sample. be done. A likelihood statistic is then computed to determine whether to accept or not to accept the candidate keyword, and therefore whether to give a final decision regarding the candidate keyword. This likelihood statistic concerns the timing of events and should not be confused with the likelihood statistic applied to multi-frame patterns with respect to the pattern of interest. In other embodiments of the invention, the range of the normalization period is set loosely, but fixed so that it cannot be changed. In this example, candidate keywords are accepted only if the normalized time period falls within fixed window boundaries. Therefore, candidate keywords are accepted only if each of the normalized times falls within the set limits. Word Level Likelihood Ratio Test Method In a preferred embodiment of the invention, each candidate word is also tested according to a likelihood ratio test method before a final decision is made to accept the keyword. This likelihood ratio test method consists of summing goodness indices for selected sequences of multi-frame patterns that have been identified as candidate keywords. The accumulated goodness index, which is the sum of the goodness indexes for each multi-frame pattern, is compared to a decision threshold value. The goodness index for a detected multi-frame pattern is the difference between the best log-likelihood statistic for any desired pattern in the keyword vocabulary and the best common score for any pattern that is allowed to be selected for the desired pattern. It is. Therefore, the goodness index has a value of zero if the best scoring target pattern is a legitimate alternative to the pattern being pursued. However, if the best score corresponds to a target pattern that is not in the list of alternatives to the selected candidate word target pattern (a given target pattern has some statistical spelling depending on accent, etc. possible), the goodness index is the difference between the best score and the best one that appeared in the list of alternatives. Decision thresholds are optimally set to obtain the best balance between missed detection and false alarm rate. Considering the word-level likelihood ratio test method from a mathematical point of view, the probability that a random multi-frame pattern x occurs is equal to p(x|k), assuming that the input speech corresponds to the target pattern class k; Probability of x". The log-likelihood statistic of the input x for the kth reference pattern is L(x|k), which is equal to lnp(x,k) as defined by equation (9). Assume that the detected multi-frame pattern must be caused by one of a group of n predetermined target pattern classes, and that these classes occur with equal frequency or that there are n possible choices. The probability of observing an event x in any case, in the sense of relative frequency of occurrence, is the sum of the probability densities defined by: For these occurrences, the proportion attributable to a given class, p(k|x), is: Or logarithmically, If the decision processor is given x and chooses class k for some reason, then equation (11A) or (11B) above gives the probability that the choice is correct. The above equation is a result of Bayes' rule. p(x,k)=p(x|k)p(k)=p(k|x)p(x) Here, p(k) is considered to be a constant 1/n. Assuming that only one class, say class m, is very promising, equation (10) can be approximated by: Furthermore, the following formula holds true. β(k,m,x)=L(x|k) −L(x|m)〓lnp(k|x) (13) If the kth class is the most promising one, then the function β is It should be noted that the maximum value is zero. When summed over a set of estimated independent multi-frame patterns, the accumulated value of β estimates the probability that the detected word is not a false alarm. Therefore, the decision threshold for the accumulated value of β is directly related to the difference between the probability of detection and false alarm and is the criterion for the likelihood ratio test method. The cumulative value of β corresponds to the index of goodness of the candidate keyword. A realized system using this speech recognition method is briefly described. As noted above, the presently preferred embodiment of the present invention provides signal and data manipulation beyond that performed by the preprocessor of FIG.
It is configured to be implemented and controlled by a Digital Equipment Corporation PDP-11 computer operating in combination with a special purpose processor such as that described in No. 841,390. A detailed program for providing the functions described in connection with the flowchart of FIG. 3 is not attached because it is not considered particularly necessary. This program printout is for digital equipment.
The PDP provided by the Corporation -
11 Macros for computers - 11 (MACRO
-11), Fortran language, and machine language for dedicated processors. From the foregoing, it can be seen that several objects of the invention have been achieved and other beneficial results have been obtained. It will be appreciated that the continuous speech recognition method described herein includes isolated speech recognition as a special application. Other applications of the continuous speech method described herein, including addition, subtraction, deletion, and other modifications of the preferred embodiments described, will be apparent to those skilled in the art and are within the scope of the following claims. It is something.

[Brief explanation of the drawing]

FIG. 1 is a flowchart illustrating the sequence of operations performed in accordance with the practice of the present invention; FIG. 2 is a flowchart of electronic equipment for performing certain preprocess operations in the overall process illustrated in FIG. 3 is a flowchart of a digital computer program implementing certain procedures in the process of FIG. 1; and FIG. 4 is a characteristic diagram illustrating classification accuracy using different conversion procedures. 13: Analog-digital converter, 17: Autocorrelator, 21: Fourier transform device, 25: Frequency response equalization, 35: Amplitude conversion, 40: Comparison, 4
2: Keyword template, 46: Post-decision processing, 51: Clock oscillator, 52: Frequency divider, 53: Latch register, 56: Digital multiplier, 58: 32 word shift register,
59: Multiplexer, 63: 32 word shift register, 65: Adder, 67: Gate, 69: 1/N division circuit, 71: Computer interrupt circuit, 7
3: Interface circuit, 85: Fast Fourier transform peripheral device.

Claims (1)

[Claims] 1. At least 1 audio signal in each of the audio signals.
characterized by a template having two target patterns and each target pattern representing at least one short-term power spectrum;
A speech analysis system for recognizing at least one predetermined keyword, comprising: collecting a pattern identified as corresponding to a sequence of target patterns; and the collected sequence of patterns is a target pattern of the keyword template. identifying candidate keywords when each corresponding to a sequence; normalizing a time interval between likelihood peaks of said collected patterns corresponding to said candidate keyword; and performing a prosodic test at said normalized time. the normalized time interval for a candidate keyword must meet a timing criterion imposed by the prosodic test before the candidate keyword is accepted as a recognized keyword. A speech recognition method that avoids 2. Applying the prosodic test comprises: applying a fixed, predetermined interval limit to each normalized interval, and wherein the normalized interval is set before the candidate word is accepted. The speech recognition method according to claim 1, wherein the speech recognition method has to fall within the limits of . 3. applying the prosodic test comprises: applying a likelihood statistic function to the normalized interval; and accepting the candidate keyword if the likelihood statistic exceeds a predetermined minimum threshold. A speech recognition method according to claim 1, comprising the steps of: 4 applying a likelihood ratio test to the sequence of collected patterns corresponding to the candidate keywords and determining a goodness index for each pattern; 2. The speech recognition method of claim 1, further comprising the steps of: accumulating the index of goodness; and accepting the candidate word if the accumulated goodness index exceeds a predetermined minimum value. 5. said step of applying a likelihood ratio test method determines the best value, referred to as the best general score, of the log-likelihood statistic for each of said collected patterns for any of said target patterns; determining the best value, referred to as the best objective score, of the log-likelihood statistic for each of said collected patterns for the objective pattern that is a valid substitute for the corresponding objective pattern of the candidate keyword; and determining a goodness index for each selected pattern by generating an arithmetic difference between the best general score and the best objective score for each collected pattern. The speech recognition method described in Section 4. 6 each of at least 1 in the audio signal
characterized by a template having two target patterns and each target pattern representing at least one short-term power spectrum;
A speech analysis system for recognizing at least one predetermined keyword, comprising: collecting a pattern identified as corresponding to a sequence of target patterns; and the collected sequence of patterns is a target pattern of the keyword template. identifying candidate keywords as each corresponds to a sequence; and applying a likelihood ratio test to the collected sequence of patterns corresponding to the candidate keywords to determine a goodness index for each pattern; A speech comprising: applying a ratio test; accumulating the goodness index for the pattern; and accepting the candidate keyword if the accumulated goodness index exceeds a predetermined minimum value. Recognition method. 7. said step of applying a likelihood ratio test method: determining a best value, referred to as a best general score, of a log-likelihood statistic for each of said collected patterns for any of said target patterns; determining the best value, referred to as the best objective score, of the log-likelihood statistic for each of said collected patterns for the objective pattern that is a valid substitute for the corresponding objective pattern of the candidate keyword; and determining a goodness index for each selected pattern by generating an arithmetic difference between the best general score and the best objective score for each collected pattern. The speech recognition method described in Section 6. 8. The audio signal is spectrally analyzed to recognize at least one predetermined keyword in the continuous audio signal, each keyword representing a plurality of temporally spaced short-term power spectra. In a speech analysis system characterized by a template having a pattern of interest, iteratively evaluates a set of parameters to determine the short-term power spectrum of the speech signal within each of a plurality of equal duration sampling periods. generating a continuous time-directed sequence of short-term audio power spectrum frames; and repeatedly selecting a first frame and at least one subsequent frame from said sequence of frames to form a multi-frame pattern. and comparing each so-formed multi-frame pattern with a respective first target pattern of each keyword template, each of the multi-frame patterns corresponding to the first target pattern of a keyword template. and, according to the determining step, selecting a subsequent short-term power spectrum for each multi-frame pattern corresponding to the first target pattern of possible candidate keywords. forming a subsequent multi-frame pattern using the selected keyword template; determining whether the subsequent multi-frame pattern corresponds to each subsequent target pattern of the possible candidate keyword templates; identifying candidate keyword templates when each of the multi-frame patterns corresponds to a target pattern of the keyword template; and normalizing a time interval between likelihood peaks of the multi-frame patterns corresponding to the candidate keywords. applying a prosodic test to the normalized time interval, wherein the normalized time interval for a candidate keyword is subjected to the prosodic test before the candidate keyword is accepted as a recognized keyword. A speech recognition method that requires meeting timing standards imposed by the above method. 9. Applying the prosodic test comprises applying a fixed, predetermined interval limit to each normalized interval, and wherein the normalized interval 9. The speech recognition method according to claim 8, wherein the speech recognition method has to be within a time limit of . 10 applying the prosodic test comprises: applying a likelihood statistic function to the normalized interval; and accepting the candidate keyword if the likelihood statistic exceeds a predetermined minimum threshold. A speech recognition method according to claim 8, comprising the steps of: 11 applying a likelihood ratio test method to the sequence of multi-frame patterns corresponding to candidate keywords and determining a goodness index for each pattern; and accumulating the goodness index for the patterns. 9. The speech recognition method of claim 8, further comprising the steps of: calculating the candidate keyword; and accepting the candidate keyword if the accumulated goodness index exceeds a predetermined minimum value. 12 said step of applying a likelihood ratio test method comprises: determining the best value, referred to as the best general score, of the log-likelihood statistic for each of said selected multi-frame patterns for any of said target patterns; determining a log-likelihood statistic, referred to as the best objective score, for each of said selected multi-frame patterns for an objective pattern that is a valid substitute for the corresponding objective pattern of the candidate keyword;
determining a best value and a goodness index for each selected multi-frame pattern by generating an arithmetic difference between a corresponding best general score and a best objective score for each selected multi-frame pattern; 12. The speech recognition method according to claim 11, comprising the step of determining.