JPS5936300A - Voice recognition equipment - 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 [Technical Field of the Invention] The present invention relates to a speech recognition device that can effectively recognize discretely uttered word sounds.

[Technical background of the invention and its problems]

When recognizing word sounds that are discretely uttered, it is very important as preprocessing to detect a sound section in which the word sound exists from an input sound signal sequence. Conventionally, however, the above-mentioned voice section has generally been detected using energy changes in the input voice. This method of detecting speech intervals has the advantage of being very simple, but in situations where the speech input environment is such that a lot of noise is added to the uttered word sounds, the above-mentioned noise It was extremely difficult to stably recognize the above word sounds by removing them. This is because if noise is added near the area where the speech to be recognized exists, it will not be possible to distinguish between the word speech and the noise using just the energy described above, and the noise will be considered part of the speech and incorporated into the recognition process. This is to put it away. Various processing methods, such as endpoint-free DP matching, have been considered to overcome these problems, but these methods have had problems such as an enormous amount of recognition processing, making them impractical.

[Purpose of the invention]

The present invention was made in consideration of these circumstances, and its purpose is to provide a simple and practical method that can stably recognize discretely uttered word sounds without being disturbed by noise. The purpose of the present invention is to provide a voice recognition device with high performance.

[Summary of the invention]

The present invention acoustically analyzes an input speech signal, extracts, for example, its energy changes and phonetic features from the acoustic analysis results, and identifies starting and ending candidate points of the speech signal according to these *'; characteristics. Detect each. Then, resampling and extracting the features necessary for recognition processing of the audio signal in a plurality of audio candidate sections obtained from all possible combinations of these starting end candidate points and end end candidate points,
Speech recognition is performed according to these sampled features.

〔Effect of the invention〕

Therefore, according to the present invention, speech recognition processing is performed using features resampled in each of a plurality of speech candidate sections, and the most reliable recognition result is extracted from the speech recognition results, including noise components. It is possible to remove the information of the detected voice candidate section and easily perform stable voice recognition there.

[Embodiments of the invention]

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of an embodiment device. Discretely uttered word sounds are acoustoelectrically converted in the audio input section 1, amplified to an appropriate signal level, and then A/D converted and captured. The acoustic analysis processing section 2 includes the above-mentioned audio input section I.
The audio signal inputted through the system is spectrally decomposed into each predetermined frame period through a dinotal bandpass filter, and its acoustic analysis is performed. However,
The feature extraction unit 3 receives the spectrally decomposed audio signal data and processes it in units of frames to extract features of the audio signal. That is, the feature extracting unit 3 extracts the phonetic feature for each frame, and for example, attaches a label indicating the vowel type to a vowel frame, and attaches a label indicating the vowel type to a nasal frame, and assigns a label indicating the vowel type to a nasal frame. For other consonants, labels such as plosive, voiced, voiceless, fricative, etc. are attached. As a result, an audio feature time series consisting of a series of labels attached to each frame of the input audio signal is obtained. In addition, the above-mentioned labeling of vowel/nasal sounds types is done by calculating the similarity between the basic spectral pattern data obtained for each frame and the standard spectral data for each type of vowel/nasal sound registered in advance in the dictionary. This is done by determining its label. Further, the above-mentioned consonant type classification is performed by examining the approximate shape of the spectral dataan obtained for each frame. In other words, if the 9 turns of the above-mentioned scuttle pattern increase monotonically along the frequency, this is determined to be frictional, and if the scuttle pattern has a so-called bell-like shape where the central part is higher, then this is judged as frictional. This is done by determining that this is rupturable.

The speech section detection circuit 4 extracts the feature time series of the speech signal consisting of the label sequence obtained by the feature extraction section 3 as described above, the spectrum data of the speech signal obtained by the acoustic analysis processing section 2, and its energy data. is input, the starting end candidate point S and the ending end candidate point E are found for the input audio signal according to this information, and the possible combinations of these starting end candidate points S and end end candidate points E are used to create the number of poles of audio candidate sections. Detected. That is, if the number "1" is now uttered as "ichi" and noise is inputted before and after the sound, the sound signal waveform becomes K as shown in FIG. 2, for example. For such an input audio signal, the starting point candidate point ""I +
Find 82 + 83 and find the end candidate point E I +F
' Find 2+E3. In the example shown in Fig. 2, the speech candidate section found from the combination of these starting and ending candidate points is as follows, since the starting point is always before the ending candidate point in time. Desired.

CS+ + Et 3. [:S+ l E2 L[Sl
+ E3: ]CS2 + Et ]tcSz + E2
, l, [Sa l E3'3 [Ss + Es:] (sz, Et) among the voice candidate sections.

As for (:s3.E3'), since its length is less than one audio frame, it may be excluded from the candidate sections to be processed.

For each of the voice candidate sections obtained in this way, the recognition unit 5 inputs, for example, spectrum information that is a characteristic of the voice signal. Then, the audio signal spectrum information sequence of each audio candidate section is resampled, its feature pattern vector is determined, and the similarity is calculated with each standard-lean vector of multiple audio categories registered in advance as a dictionary. Then, voice recognition processing is performed. Regarding the plurality of speech candidate sections obtained as described above, the recognition unit 5
Each of the above recognition processes is performed and the recognition results are sent to the control unit 6.
He is appearing in This control unit 6 includes each of the processing units 2 described above.
, 3°4.5, and input the recognition results obtained by the recognition section 5 for each speech candidate section to make a comprehensive judgment. Then, if the speech candidate section includes noise, this noise naturally deteriorates the recognition result (similarity value), so this is removed, the most reliable recognition result is extracted, and the This is being presented as a correct recognition result for audio signals. Thus here,
Recognition results obtained from highly reliable speech candidate sections are obtained, and recognition results obtained from information containing noise added before and after the speech are effectively eliminated. Only the information in the speech candidate sections that do not contain h1 noise among the number of speech candidate sections is effectively extracted.
, will be recognized.

Note that the speech recognition process for each speech candidate section in the recognition section 5 may be performed using various conventionally proposed methods as appropriate. The characteristics of the voice used in Katakotako's recognition process are also
Needless to say, various methods can be adopted.

By the way, the main feature of this device is the detection of the starting point candidate point S and the ending point candidate point E for the audio signal, and the detection process of the audio candidate section that can be determined from the combination of these starting point and ending point candidate points S and E. is the voice section extraction circuit 4
This is done by the VC as follows. FIG. 3 is a flowchart showing an example of the processing process. This processing is performed by first initializing the processing control color value and then inputting the audio signal energy of the nth frame. After that, for example, divide the audio signal into a silence class and a voice class according to the temporarily set threshold, find the edge variance between each class, set the optimal threshold Eth that maximizes the value, and calculate the threshold Eth and the input audio. The energy E(n) is compared. Thereafter, the time point at which the input audio energy E(n) exceeds the threshold value FEth is defined as S'(1) and is set as the first candidate point for the start/end. Then, the point in time when the input audio energy E(n) falls below the threshold Eth is detected, and this is set as the first candidate point E'(1) of the termination point. Then, find the interval between the starting and ending candidate points found in this way as Tk '' I S'(lc) E'(k) l, and determine whether it exceeds 1♀6 T'th while Hr is constant. The speech candidate section is determined by determining whether or not the speech candidate section is the same.As a result, the erroneous speech family ++!i section that is obtained fragmentarily is removed. All frames are sequentially input, all possible combinations of 5'f are checked, and all voice candidate sections are obtained.

Thereafter, the speech characteristics of the speech candidate section are examined, it is determined whether or not the section contains a noisy component, and the 11-minute complementary section of speech containing the noisy component is removed from the recognition target.

After that, the end of the utterance is detected, for example, from the interval E(n)<Eth or n(which continues for a certain period Mth, and this
The final speech candidate section is determined by comparing the speech features in the speech candidate section detected so far with the utterance shape of the word registered in advance. For all of the speech candidate sections determined in this way, for example, the composite similarity method is used to calculate the degree of similarity with the word dictionary registered in the dictionary, and the similarity values are compared with each other to find the most reliable one. Outputs highly accurate results as recognition results.

The recognition process described above compares the t3 identification results obtained for each candidate section containing the word sound and extracts the most reliable one, so the final utterance is discretely uttered. The recognition results obtained from the characteristics of the voice itself can be obtained with high reliability. Since the detection of the distorted speech section and the recognition process are performed in conjunction with each other, a great practical effect is achieved in that the recognition process can be performed stably.

Note that the present invention is not limited to the above embodiments. For example, various methods can be employed as appropriate for the process of detecting the start end candidate point and the end candidate point, the process of extracting the voice candidate section obtained by a combination of these, and the recognition process.

There are no particular limitations on the characteristics of the audio used in the above-described processing. In other words, the present invention can be modified in various ways without departing from the spirit thereof.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram of a device according to an embodiment of the present invention, FIG. 2 is a diagram showing an audio signal waveform, start and end candidate points, and its audio candidate section, which shows the processing of the embodiment device. FIG. 2 is a diagram showing the flow of recognition processing in the embodiment device. ! ...Audio input section, 2...Acoustic analysis processing section, 3...
・Feature extraction unit, 4... Voice section detection circuit, 5... Recognition unit, 6... Control unit.

Claims (1)

[Claims] means for inputting and acoustically analyzing an audio signal; means for extracting features of the audio signal from the acoustic analysis results; and determining start and end candidate points of the audio signal according to the features and the acoustic analysis results. means for respectively detecting, means for resampling and extracting the features of the audio signal in a plurality of audio candidate sections obtained by all possible combinations of the starting end candidate point and the ending end candidate point, and the resampled features of these resampled features. A speech recognition device comprising: means for recognizing speech included in the speech signal using the speech recognition apparatus.