JPS6011898A - 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 highly practical speech recognition device that can recognize phonemes, words, etc. of input speech with high accuracy.

[Technical background of the invention]

Recently, with the advancement of electronic technology, various speech recognition devices have been developed. For example, in word speech recognition targeting a specific speaker, a considerably high recognition rate can be obtained even if the recognition target is only a few six words. It has become. Furthermore, even in word speech recognition for unspecified speakers, it has become possible to recognize several dozen words.

However, when dog language words or continuously uttered speech are to be recognized, it is difficult to learn word by word or to change the history of the vocabulary to be recognized. For this reason, it is thought that a method should be adopted that first performs recognition processing at the syllable and phoneme level, which are the constituent elements of input speech, and then performs recognition processing using linguistic information and a word dictionary. If such a method is adopted, by using phonemes as the basic unit of recognition, it will be possible to recognize not only word sounds but also continuously uttered sounds by using the results of approximately 20 types of phoneme recognition. Are known. In such cases, recognition performance strongly depends on the phoneme recognition rate, so how to perform the above-mentioned speech analysis and phoneme recognition accurately and efficiently is an important issue. It has become.

[Problems with background technology]

By the way, phonemes (phonemes), which are the constituent elements of speech, can be classified into various types depending on their vocalization mechanisms. That is, phonemes are first broadly classified into vowels and consonants. The consonants are classified according to voiced/unvoiced, articulatory position, articulatory form, etc. Among the above-mentioned phonemes, vowels have the property of being stationary and having a long duration.

Another characteristic is that the duration of consonants is short.

For example, voiceless plosives (p, t, k) have rapid voice changes and short durations. And its characteristics have the property of appearing at the point of rupture.

However, in conventional speech recognition, input speech is analyzed for a certain analysis time length and at a certain analysis frame period, without considering the properties of each phoneme, and the characteristics obtained by this are analyzed. Recognition processing is performed using parameters.

For this reason, when recognizing vowels, the analysis time length is short, so spectrum analysis using discrete Fourier transform etc. has poor frequency resolution, and the analysis results become unstable due to the influence of the pitch frequency of the vocal cord sound source. There was a problem in that erroneous recognition was likely to occur. On the other hand, regarding consonants that change rapidly such as the voiceless plosives (p, t, k) mentioned above, the analysis time length and analysis frame length are too long, so the above-mentioned There was a problem where important information about consonant changes was buried under other information.

There is also a word speech recognition method that uses tscutane matching for the entire word without performing phoneme recognition, but as the number of words to be recognized increases, the number of similar words in the recognition target increases. There was a problem that identification became difficult. In addition, even in such a recognition method, the feature parameters of the input speech are extracted by keeping the analysis time length and analysis frame period constant, so there may be cases where the word speech feature vector is unstable or there is a noticeable difference. There was a problem that it was difficult to see the characteristics.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to easily and accurately identify phonemes, words, etc. by effectively utilizing the characteristics of the phonemes that make up input speech. An object of the present invention is to provide a highly practical speech recognition device that can recognize speech.

[Summary of the invention]

The present invention calculates the time series of the characteristic A' parameter of input speech for each of a plurality of analysis times with different analysis time lengths and analysis frame periods, and extracts a part of the time series of these feature parameters as a speech feature vector. This method performs speech recognition by pattern matching.

That is, when audio analysis is performed on input audio using, for example, digital signal processing, a plurality of analysis times (T+, F
, )(T21F2)(T3°Fa)~, digital bandpass filter group, discrete 7-lier transform (DFT),
The time series of the characteristic parameters of the input speech X 1 +
2 HX 3～, the time series of these characteristic parameters
1 to (i=1°2.3 to N) are extracted as their speech feature vectors, and similarity, such as distance and likelihood, is calculated between them and a speech category dictionary prepared in advance to identify phonemes, syllables, It is designed to recognize input speech such as words and sentences. Same, the time series of the feature/4' rammeter X! +Xt
It goes without saying that +X3~ can be modified by analog processing.

〔Effect of the invention〕

Thus, according to the present invention, since the time series of the feature vectors of the input speech are examined for each of a plurality of analysis times with different analysis time lengths and analysis frame periods, relatively stable and stationary features such as vowels are Regarding phoneme recognition, the time series of feature parameters requested through analysis processing with a longer analysis time length and analysis frame period is used as a feature vector, making it possible to recognize this effectively and accurately. . In addition, for phonemes such as plosive consonants that change quickly and have non-stationary properties, the time series of feature parameters determined by analysis with short analysis time length and analysis frame period can be used as feature vectors. It becomes possible to recognize easily and with high precision. In this way, by pattern matching the information on each recognized phoneme, that is, the time series/turns of the phoneme and the phonetic category dictionary, the input speech can be effectively processed. It becomes possible to recognize the

Furthermore, as described above, since the time series of the characteristic parameters of the input speech is obtained by varying the analysis time, it is possible to improve the accuracy of vowel analysis and perform recognition with emphasis on vowels. Furthermore, since the features of consonants can be well extracted, it is easy to distinguish between similar words with different consonants. Therefore, by effectively utilizing a plurality of types of feature vectors obtained at different analysis times, simple and highly accurate recognition can be performed, which is extremely effective in practice. In addition, highly accurate speech recognition that effectively utilizes the properties of phonemes that make up speech becomes possible.

[Embodiments of the invention]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a schematic diagram of an embodiment of an apparatus in which phonemes are the basic unit of recognition processing. The input audio is passed through first and second analysis circuits 1, which have different analysis time lengths and analysis frame cycles.
2, and are analyzed at each predetermined analysis time to extract characteristic parameters. These analysis circuits 1 and 2 are configured as shown in FIGS. 2 to 5, which will be described later. A time series of characteristic parameters of the input voice is obtained by performing predictive analysis processing, analog band filter processing, etc. The time series of feature parameters determined by these analysis circuits 1 and 2 are temporarily stored in a phoneme feature vector memory, and then a part of them is read out as a feature vector and sent to similarity calculation circuits 3 and 4. each is given. The similarity calculation circuit 3 calculates, for example, the analysis time length of 20 m5 and the frame period of 10 determined by the analysis circuit l.
From the feature vector in a long analysis time consisting of ms,
Phonemes with constant and stable properties, such as vowels, are recognized by checking with the phoneme dictionary 5 and calculating the degree of similarity. The other similarity factor η-circuit 4 uses the feature vectors determined by the analysis circuit 2 for each short analysis time, for example, where the analysis time length is ]OmS and the frame period is 5 ms, to be compared with the phoneme dictionary 6. Yosi, tsu'-! Rapidly changing phonemes such as consonants are recognized by calculating similarity. The time series of phoneme recognition results obtained by these similarity calculation circuits 3.4, that is, the vowel phoneme pattern and the consonant phoneme/vowel turn, are, for example, a string of phoneme symbols including the similarity value. are stored in the phoneme pattern memory 7 and synthesized.

After that, the phoneme pattern for the input speech obtained in the phoneme pattern memory 7 is given to the matching circuit 8, and the similarity is calculated with the speech category dictionary 9, which is formed by registering standard phoneme patterns of words, for example. It has become recognized.

That is, in this device, the analysis time length T and its frame period F in the analysis circuit 1 are determined by the time series of the characteristic parameters of the input speech determined over a long analysis time, so that the input speech has a stationary property. Phonemes such as vowels are effectively recognized.

In addition, in the analysis circuit 2 in which the analysis time length T and frame period F are set short, phonemes such as consonants that change quickly in the input speech are It becomes possible to effectively reduce the Since the entire input speech is recognized using the phoneme pattern made up of the time series of the phonemes of these vowels and consonants, the recognition accuracy can be made sufficiently high. Furthermore, the recognition processing described above is simple and therefore has great practical advantages.

Now, the above-mentioned analysis circuits 1 and 2 can be configured as follows, for example. FIG. 2 is a configuration diagram of an analysis circuit that extracts characteristic parameters by frequency analysis using a group of digital bandpass filters. This circuit converts the input audio into A/
A Kr digital signal is converted via a D converter 11 every 100 μs, for example, and inputted to a plurality of band pass filters (DBPF) 12 provided in parallel to separate it into frequency components of 16 channels. Thereafter, the outputs of each of these channels are input to the 2-east circuit 13 to calculate the power thereof. Thereafter, these power components are passed through a first low pass filter (LPF) 74 group and a second LPF 15 group.
The phoneme features A'
It is configured such that the parameters are collected and stored in the memo IJ I 6 , 17.

Here, the first LPF 14 group has an analysis time length (M minute time) of 24 ms and extracts feature parameters xI every frame period of 8 ms, and the second LP
In the F15 group, the analysis time length is 8 ms, and the feature/crameter y1 is extracted at a frame period of 8m54H. The time series of the feature parameters x1 determined in this way is used for recognizing phonemes such as vowels, and the time series of the feature parameters yI is used for recognizing phonemes such as consonants.

According to the analysis circuit having such a configuration, the configuration of the bandpass filter 12 is the same or is used in common, and the LP
By simply changing the smoothed frame length (integration time) of F 14 and J 5, feature parameters Xl with different time lengths can be obtained.
r ''/i.Moreover, in this case, the amount of calculation for analysis processing does not increase too much, it is economical in terms of circuit configuration, and it is advantageous in practice.
It is also effective to separately determine the frame period of F14°15 as 8ms and 4ms.

Furthermore, when the analysis circuit 1.2 is configured using a group of analog bandpass filters, it may be configured as shown in FIG. In other words, input audio is passed through a 16-channel analog bandpass filter (
BPF),? After passing through the 1st group, input to the 2nd east circuit 22nd group to increase power. After that, the output wave is converted into f-waves through analog low-pass filters LPF 23 and 24, and then digitally converted through A/D converters 25 and 26, and its characteristics are stored in memory Y6.17. rixI,y
All you have to do is set K to store l. In this case, the LPF
23 is set to a long time constant of 30 ms, for example, and digitally converted via an A/D converter 25 with a sampling period of 15 m5 to obtain characteristics (characteristics suitable for recognizing phonemes such as vowels) and parameters Xi. Do it like this. Then, the time constant of LPF24 is set as short as, for example, 10 mS,
By setting the sampling period of the A/D converter 26 to, for example, 5 mS, characteristics and parameters suitable for recognizing phonemes such as consonants may be obtained.

Even if the analysis circuit is configured in this way, the BPF 21 can be shared, so it is possible to improve the feature parameter extraction performance without significantly increasing the size of the circuit. It is also economical and efficient. However, this embodiment can also be implemented using a swinoted capacitor filter.

Further, FIG. 4 shows an example of an analysis circuit configured by applying kegstral analysis processing.

In this case, input audio input manually via the A/D converter 11 is divided into two systems, and a time window circuit with a Hanning window of 24 ms, for example, is used. 1 and a time window circuit 32 provided with a Hanning window of 12 m5. Thereafter, the signals extracted in each of these time windows are input to a discrete Fourier transformer (DFT) 33.34 VC for spectrum conversion, and the output thereof is input to an absolute value circuit 35.34 VC. ．． 36 and logarithmic transform circuits (LOG) 37 and 38, and then undergoes cepstrum transformation in discrete inverse Fourier transformers (IDFT) 39 and 40. As a result, the cepstrum for the signal with a time window of 24 mS is obtained as the output of the IDF'r 39, and the top 16 quesonutramnosorameters are extracted as phoneme feature parameters ei.

The IDE'T 40 also obtains the cepstrum for the signal in the time window 12m5, and its top 16 cepstrum parameters are taken as phoneme feature parameters c'i. Then, these are stored in the memories 16 and 17 respectively, and the time series of the feature parameters ei is used for phoneme recognition such as vowels.
Further, the time series of the feature parameters C'i is used for phoneme recognition such as consonants.

In this way, by changing the analysis time window and performing cepstrum processing, we can obtain a time series of the feature parameter c1 that reflects the characteristics of the vowel, and a time series of the feature parameter alt that reflects the characteristics of the consonant. be able to.

Furthermore, FIG. 5 shows a modification of the analysis circuit using the digital bandpass filter group described above.F), LPF
A predetermined number of the outputs of 15 are input to the adder (ADD) 1B, and the sum is calculated to obtain the feature parameter yt of the 4 channels.
This is stored in the memory 19. That is, in this example, the number of frames is set to "1" and the number of frames is set to 16th order in the frequency direction as characteristic parameters for recognizing stationary phonemes such as vowels.

In addition, as features and parameters for recognizing rapidly changing phonemes such as consonants, the number of frames is increased to 4 frames, and the frequency direction is set to 4th order. In this way, even if the frequency resolution of the feature/yarameter for consonant recognition is degraded, the feature vector is busy with 47 frames and is sensitive to changes, so it is not sufficient to recognize consonants etc. It is possible to provide Furthermore, although the characteristic parameters for vowel recognition cannot be determined for one frame, since their nature is stationary, the purpose can be achieved by simply making the frequency resolution sufficiently high. Therefore, the effect is that the configuration can be significantly simplified compared to the circuit shown in FIG. 2. In addition, as a modified example of this embodiment, L
It is also possible to add the outputs of the LPF 15 for a plurality of frames and transfer the result to the memory 16 without using the PF 14.

In this way, the analysis circuit 1.2 can be configured relatively easily by sharing a part thereof. The analysis time length T, frame period F, etc. in this case are based on the LP described above.
This can be easily changed by changing the time constant of F, the length of the time window, the sampling rate of the A/D converter, etc. And what about the characteristics depending on the nature of the phoneme? parameters can be obtained easily and effectively.

Although an embodiment in which phonemes are used as the basic unit of recognition processing has been described above, the present invention can also be implemented as a speech recognition device that performs word-by-word matching. FIG. 6 is a schematic diagram of the configuration, and the analysis circuits 1 and 2 are composed of a group of 16-channel digital bandpass filters as shown in FIG. In this case, the analysis circuit 1 obtains a steady time series of characteristic parameters through analysis processing with an analysis time length of 30 mS and a frame period of 15 mS, for example. In addition, in analysis circuit 2, the analysis time length is 10m5. Through analysis processing with a frame period of 5 mS, the time series of features and parameters that are sensitive to changes in speech can be determined. A resampling circuit 41 is applied to the time series of such feature parameters.
42, resampling is performed at equal intervals from the start frame to the end frame of the voice determined from the power information of the input voice, for example, 16 dimensions in the frequency direction and 16 dimensions in the time axis direction.
Dimensional Tears: 256 dimensions of time, frequency, and pattern. These time/frequency patterns are stored in word feature vector memos IJ 43 and 44, respectively.

These time/frequency patterns differ depending on the analysis time length and frame period during each analysis process. The pattern obtained in the vector memory 43 has poor resolution in the time axis direction compared to the pattern obtained in the vector memory 44, resulting in a so-called blurred pattern, but on the contrary, it is stable. There is. The similarity calculation circuits 45 and 46 use word dictionaries 47.
48 to determine the word similarity.

Note that as the similarity calculation method used here, it is useful to use a composite similarity method or a statistical measure such as Mahalanobis distance.

Thereafter, the word is comprehensively evaluated and recognized by the comprehensive judgment circuit 49 using the similarity value for each word category determined using these different feature vectors and the similarity difference. . Similarly, in this case, the similarity result obtained based on the stable word feature vector obtained by increasing the analysis time length and frame time length is used for the general classification process for word categories, and the analysis time length and frame time length are It is desirable to use the similarity results obtained based on word feature vectors with shortened time lengths to identify similar words;;); :tr. Furthermore, the multiple types of matching processing in this embodiment can also be performed hierarchically.

As the analysis circuit 1.2 used for this word matching, an analysis circuit using an analog bandpass filter, Cepnutrum analysis, discrete Fourier transform, etc. as shown in FIGS. 3 to 5 described above may be adopted. Of course it is possible.

As described above, according to the present invention, since speech is recognized by sufficiently effectively utilizing the characteristics of the phonemes constituting the speech, highly accurate recognition results can be obtained. Moreover, recognition results can be obtained effectively through simple processing.
Its practical advantages are enormous.

Similarly, the present invention is not limited to the above embodiments. For example, the analysis time length and frame period may be determined according to the specifications of the input audio, and the number of analysis times determined by the analysis time length and frame period may be plural.

In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram of one embodiment of the present invention, FIGS. 2 to 5 are diagrams each showing a configuration example of an analysis circuit, and FIG. 6 is a schematic configuration diagram showing another embodiment of the present invention. be. 1.2...Analysis circuit, 3.4...Similarity calculation circuit,
5.6... Phoneme dictionary, 7... Phoneme/zo turn memory, 8... Verification circuit, 9... Speech category dictionary, 11
...A/D converter, 12...DBPF', 13...
・Squaring circuit, 14°15...LPF, 16, 17.1
9... Feature vector memory, 2... BPF, 22
...Squaring circuit, 23°24...LPF, 25, 2
6... A/D converter, 31°32... Time window circuit,
33.34...DFT, 35°36...Absolute value circuit, 37.38...Logarithmic conversion circuit, 39.40...Bow D
FT, 41, 42...Resample circuit, 43.4
4...Word feature vector memory, 45°46...Similarity calculation circuit, 47.48...Word dictionary, 49...
Comprehensive judgment circuit.

Claims (2)

[Claims] (1) Means for analyzing input speech for each of a plurality of analysis times determined by varying the analysis time length and analysis frame period to determine a time series of characteristic parameters of the input voice, and these characteristic parameters. 1. A speech recognition device comprising means for extracting a part of the time series as a speech feature vector and comparing the speech feature vector with a pre-registered speech category dictionary. (2) The means to obtain the time series of the characteristic nomurameter of the input speech is to apply different sampling processes to the same intermediate results obtained in a series of analysis processing processes for analyzing the input speech, and to determine the analysis time length and the analysis frame period. 2. The speech recognition device according to claim 1, wherein time series of feature parameters are obtained for different analysis times.