JPH0455518B2 - - Google Patents

Description

[Detailed description of the invention]

FIELD OF THE INVENTION The present invention relates to a speech recognition device for automatically recognizing the content of speech. Conventional structure and its problems In speech recognition targeting unspecified speakers,
The characteristics of voices vary greatly depending on gender or age, and the challenge is how to standardize the characteristics of voices and recognize the voices of unspecified speakers. When recognizing speech phoneme by phoneme, standard phoneme patterns vary greatly depending on gender and age; for example, for the vowel /a/, there is a large difference in spectral shape between men and women. In order to deal with this problem, conventionally, multiple standard patterns are prepared for the same phoneme, and the similarity of all standard patterns is calculated for the input speech to determine which standard pattern is most similar. It is then recognized. However, this method has the drawback that the larger the number of standard patterns to be prepared, the more likely they are to be confused with each other, resulting in lower recognition performance and an enormous amount of calculation. An example of a block diagram of a conventional speech recognition device is shown in the first example.
As shown in the figure. First, the speech data of multiple speakers is divided into groups using a clustering method or the like, and standard pattern groups are created in units of phonemes or syllables and stored in the standard pattern storage section 11. For the sake of explanation, it is assumed here that standard pattern group 1 in standard pattern storage 11 is data for only men, and standard pattern group 2 is data for women only, and six types of standard patterns are prepared for each group. Now, the input voice inputted from the microphone 1 is subjected to AD conversion by the AD converter 2, and then one side is sent to the signal processing circuit 3, which performs pre-emphasis and window calculation, and then is sent to the linear predictive analysis processor 4. A.D.
The other converted signal is sent to the segmentation unit 5
Here, band power calculation, voice interval detection, voiced/unvoiced/silent determination, and consonant voice segmentation are performed, and the results are transferred to the main memory 7. Using the LPC parameters obtained by the linear predictive analysis processor 4, the similarity calculation section 6 performs similarity calculation according to the following procedure. First, the standard patterns in the standard pattern group 1 stored in the standard pattern storage section 11 are transferred to the similarity calculation section 6, where similarity calculation is performed for each frame and transferred to the main memory 7. The same process is performed for standard pattern group 2 as well. The main processor 8 refers to the main memory 7, adopts as a recognition result the phoneme or syllable corresponding to the standard pattern with the highest degree of similarity for each frame, and uses this and the result of the segmentation unit 5 to identify the phoneme or syllable. Create a sequence of syllables. Word recognition is then performed by comparing the resulting series with the word dictionary 12, and the results are sent to the output section 9. The disadvantage of this conventional example is that similarity calculations must always be performed for all the standard pattern groups stored in the standard pattern storage unit 11.1 The amount of calculation in the similarity calculation unit 6 becomes large, and high-speed calculation is required. The required cost is high. 2. Since this method prepares multiple standard pattern groups and uses the phoneme with the highest degree of similarity among all the standard pattern groups for recognition, the number of similar phonemes increases, leading to confusion with each other. This increases the recognition performance. Purpose of the Invention The present invention solves the above-mentioned drawbacks and automatically selects a standard pattern that is most suitable for the unknown input speech. It is an object of the present invention to provide a speech recognition device that enables recognition with high reliability and can perform high-speed processing by significantly reducing the amount of calculation for similarity calculation. Structure of the Invention In order to achieve the above object, the present invention divides the voices of many speakers into groups in advance and obtains a standard pattern of the voice for each group, and when an unknown voice is input, the standard pattern of the voice is The degree of similarity is calculated, and the degree of confidence that the unknown voice belongs to that group is calculated for each group using the degree of similarity. When the degree of confidence exceeds a certain threshold, only the standard patterns of the group that exceeded the threshold are selected. The present invention relates to a speech recognition device characterized in that it is used to recognize unknown input speech thereafter. DESCRIPTION OF EMBODIMENTS The present invention is characterized in that the voice of an unspecified speaker can be stably recognized regardless of age or gender. To this end, standard patterns for recognizing voices are created by dividing them into groups according to the quality of the voice, such as male, female, child, elderly, etc.
It is unknown which group the input voice actually belongs to, but by using this method, the input voice can be analyzed and automatically determined with high accuracy which group it belongs to. can be determined. In the following embodiment, a case will be described in which only two groups, men (group 1) and women (group 2) are targeted. FIG. 2 shows a configuration diagram of the speech recognition device of the embodiment. First, the contents stored in the standard pattern storage section 25 will be explained. In this example, the average value is determined for each group 1 and 2, the covariance matrix is determined for the entire group, and these are used to calculate the weighting coefficient, which is a standard pattern.
a _ij and average distance d _i are determined and stored in the standard pattern storage section 25. First, LPC of voice i in group 1 voice
Let the average value of the cepstral coefficients be m _i ⁽¹⁾ = (m _i1 ⁽¹⁾ , m _i2 ⁽¹⁾ , ......, m _ip ⁽¹⁾ ). In the formula, (1) indicates group 1,
p is the number of parameters used. When the standard pattern is constructed on a frame-by-frame basis, if the order of the LPC cepstrum coefficient is p, then P=p. If the standard pattern consists of a time pattern of n frames, use P.
=p×n. Also, the LPC of phoneme i in group 2 speech
Let the average value of the cepstral coefficients be m _i ⁽²⁾ = (m _i1 ⁽²⁾ , m _i2 ⁽²⁾ , ..., m _ip ⁽²⁾ ). This is applied to the vowels /a/, /i/, /u/, /e/, / for each of Group 1 and Group 2.
Obtain o/ and nasal sounds, making a total of 12. Next, using the average value m _i ⁽¹⁾ of group 1 and the average value m _i ⁽²⁾ of group 2, let the covariance matrix common to these 12 phonemes be R, and its inverse matrix be R ^-1 . . If the (j, j') element of R ^-1 is r _jj ', then the weighting coefficient for the jth order of the LPC cepstrum coefficient is a _ij ⁽¹⁾ = Z _P 〓 ^j=1 r for phoneme i of group 1. _jj ′m _ij ⁽¹⁾ …(1). Further, if the average distance to voice i is d _i ⁽¹⁾ , then it is calculated as d _i ⁽¹⁾ = m _i ^(1)t R ^-1 m _i ⁽¹⁾ ...(2). (t represents a transposed matrix.) These a _ij ⁽¹⁾ and d _i ⁽¹⁾ are obtained for each phoneme and stored in the standard pattern group 1 of the standard pattern storage section 25. Similarly, find a _ij ⁽²⁾ and d _i ⁽²⁾ for group 2,
Store in standard pattern group 2. It goes without saying that the covariance matrix may be obtained for each group 1 and 2. Next, the operation of the similarity calculation section 24 will be explained. When an unknown input voice is input from the microphone 20 (assuming that it is the voice "ha zi me"), the signal processing circuit 22 performs pre-emphasis and window calculation, and then the linear predictive analysis processor 23 inputs the unknown input voice. and add a coefficient c _j (j=
1, 2, ..., p) are found. When using a time pattern, line up c ₁ to c _p of (n frames) and use c ₁
The similarity calculation unit 24 calculates the similarity between c _p and the standard pattern sent through the standard pattern switching unit 26 and c _j . For the phoneme i of group 1, the similarity l _i ⁽¹⁾ is obtained as l _i ⁽¹⁾ = _P 〓 ^j=1 a _ij ⁽¹⁾ c _j dm _i ⁽¹⁾ ...(3). This is also obtained for group 2 using l _i ⁽²⁾ = _P 〓 ^j=1 a _ij ⁽²⁾ c _j −d _i ⁽²⁾ ...(4), and a total of 12 phonemes are obtained and stored in the main memory 27. Forward. The segmentation unit 26 calculates the band power and voiced/unvoiced judgment, determines the speech interval, and detects the consonant interval (here, /h/ of ha zi me, /
z/, /m/) and transfer it to the main memory 27. The main processor 28 has main memory 2
Vowels using the consonant intervals and similarity registered in 7.
Determine the nasal interval (here / of ha zi me)
a/, /i/, /m/, /e/), N phoneme centers (center position or position with maximum similarity) are found (here, N=4). Next, the operation of the selection section 29 will be explained. First, the four phonemes (/a/, /
Prepared all phonemes (/a/, /i/, /u/, /) at the phoneme center of i/, /m/, /e/)
The maximum similarity for each group (e/, /o/, nasal) is determined for each group, and the case of group 1 is set as l _i ⁽¹⁾ , and the case of group 2 is set as l _i ⁽²⁾ . Obtain this for each N phoneme center, and calculate the sum of similarities for each group.
Let L ⁽¹⁾ and L ⁽²⁾ . L ⁽¹⁾ = _N 〓 ⁿ⁼¹ l _i ⁽¹⁾ ……(4) L ⁽²⁾ = _N 〓 ⁿ⁼¹ l _j ⁽²⁾ ……(5) This L ⁽¹⁾ , L ⁽²⁾ The reliability R _e is defined using the following formula. R _e ⁽¹⁾ L ⁽¹⁾ −L ⁽²⁾ ……(6) If there are three or more groups, calculate the sum of the similarities for each group, and calculate the two maximum similarities using the above (6). ) can be used to find the reliability R _e ⁽¹⁾ . Now, if this R _e ⁽¹⁾ is a positive value and exceeds a predetermined threshold, it is determined that the user's voice belongs to group 1. If it is a negative value and its absolute value exceeds the threshold, it is determined that the user's voice belongs to group 2. After the determination is made by exceeding the threshold value, the selection unit 29 instructs the standard pattern switching unit 32 to provide only the standard patterns of the determined group to the similarity calculation unit 24, and performs the operation. end. If Re does not exceed the threshold, the selection section 29
instructs the standard pattern switching unit 32 to select the standard patterns of both groups 1 and 2, and also instructs the main processor 28 to select the similarity of group 1 if the value is positive, or to select the similarity of group 1 if the value is negative. gives an instruction to use the similarity of group 2 for phoneme recognition. Therefore, the main processor 28 follows the instructions of the selection unit 29 as long as the reliability Re does not exceed the threshold, performs phoneme recognition using the instructed similarity, and compares the results with the word dictionary 30 to identify the word. Recognition is performed, and the word dictionary with the highest degree of similarity is transferred to the output unit 31 as a recognition result. Further, while the reliability Re does not exceed the threshold, the standard pattern switching section 26 sequentially transfers the standard patterns 1 and 2 according to the instructions of the selection section 29, and the similarity calculation section 24 transfers the similarity to the standard pattern groups 1 and 2. Repeat the calculation. Therefore, during this time, the similarity calculation unit 24 requires a large amount of calculation for similarity calculation, but from the time when the operation of the selection unit 29 is finished, it is only necessary to calculate the similarity of the determined group, and the calculation amount is reduced. is significantly reduced. Furthermore, the main processor 28 can now perform phoneme recognition using highly reliable standard patterns, improving the accuracy of word recognition. In the embodiments described above, the reliability is calculated using the sum of the maximum similarities, but the reliability may also be calculated using the number of times the maximum similarity is obtained. Among the similarities obtained by the similarity calculation unit 24,
Let l _i ⁽¹⁾ be the one with the highest degree of similarity. The main processor 28 notifies the selection unit 29 that l _i ⁽¹⁾ . The selection unit 29 assumes that l _i ⁽¹⁾ belongs to group 1 and counts up the number of times N ⁽¹⁾ . l _j
If ⁽²⁾ is sent, count up the number of times N ⁽²⁾ . Reliability Re is calculated using the following formula. Re ⁽¹⁾ =N ⁽¹⁾ /N ⁽²⁾ ...(7) Re ⁽²⁾ =N ⁽²⁾ /N ⁽¹⁾ ...(8) Which of Re ⁽¹⁾ and Re ⁽²⁾ If the value exceeds a predetermined threshold, group 1 is determined if it is Re ⁽¹⁾ , and group 2 is determined if it is Re ⁽²⁾ .
As long as the threshold is not exceeded, N ⁽¹⁾ and N ⁽²⁾ are compared, and the similarity of the larger group is used for phoneme recognition.
An instruction is given to the main processor 28. Since this method uses only the number of times to obtain the maximum similarity, it has the advantage of being more stable against factors that distort the speech spectrum, such as noise, compared to the method using the sum of similarities described above. Note that when there are three or more groups, the reliability Re can be calculated for the two groups that have the highest number of times. Next, automatic selection of a standard pattern group will be explained using the block diagram of FIG. 2 and the flowchart of FIG. 3. As shown in Process A, suppose that an arbitrary word, for example, the voice "ha zi me" is input into the microphone. The audio is A/D converted by the A/D converter 21 (processing 2), and one is sent to the signal processing circuit 22 and the other is sent to the segmentation section 26. The signal processing circuit 22 performs pre-emphasis and window calculation using a Hamming window for each frame, as shown in FIG.
send to The linear predictive analysis processor 23 performs linear predictive analysis to obtain LPC cepstral coefficients C=(c ₁ ,
c ₂ , . . . c _{j ,} . On the other hand, the segmentation unit 26 performs band filter calculations (processing E), and uses the LPC cepstrum coefficients C obtained by the linear predictive analysis processor 23 to determine voicedness and unvoicedness, detect voice sections (processing E),
Segmentation and consonant discrimination (processing) of the consonants /h/, /z/, /m/ of the consonants ha zi me are performed and the results are sent to the main memory 27 (processing). The selection unit 29 also sends a _ij ⁽¹⁾ and d _i ⁽¹⁾ prepared in advance in the standard pattern group 1 to the similarity calculation unit 24 (processing). The similarity calculation unit 24 calculates the similarity l _i for the phoneme i of group 1 using the following equation as shown in Process
Find ⁽¹⁾ . l _i ⁽¹⁾ = _P 〓 ^j=1 a _ij ⁽¹⁾ c _j −d _i ⁽¹⁾ ...(3) The similarity is preferably based on a statistical distance measure such as Bayesian judgment or Mahalanobis distance. . Similarly, l _i ⁽²⁾ is obtained for group 2, and these are transferred to the main memory 27 (processing). The main processor 28 refers to the segmentation result and the degree of similarity for vowels and nasals to process the vowel/nasal parts ha zi me /a/ and /.
Determine i/, /m/ and /e/, and determine the determined vowel,
From among the nasal parts, the center frame (center position or position with maximum similarity) that is most likely to be a vowel or nasal sound is selected for each vowel/nasal part, and its position information is provided to the selection unit 29. The selection unit 29 determines the maximum similarity for each group of central frames, and further determines the sum of the similarities L ⁽¹⁾ and L ⁽²⁾ . Then, the reliability is calculated using equation (6) or equations (7) and (8), and it is determined whether the threshold value is exceeded or there is distortion (processing).
Based on this result, the standard pattern switching unit 32 selects a standard pattern group in the standard pattern storage unit 25. Next, FIG. 4 shows the flow of processing of the speech recognition apparatus according to this embodiment. When speech is first input (judgment 2), it is acoustically analyzed (processing), and then segmentation and similarity calculation 1 are performed via the judgment (processing). At this time, similarity calculation is performed for standard patterns of all prepared groups. Next, the phoneme centers of vowels and nasal sounds in the speech are extracted, and the reliability of group discrimination is calculated (processing step). If the reliability is below the threshold (judgment c), phoneme recognition is performed using the similarity of the group with the highest reliability at that point. If it exceeds the threshold, a standard pattern selection end command is issued (processing ma), and phoneme recognition is performed based on the similarity of groups exceeding the threshold (processing a). Word recognition is performed using the phoneme recognition results (processing wo), the word recognition results are output (processing wa), and the process returns to the voice input waiting state. When the next voice is input, after acoustic analysis (processing), check whether a standard pattern selection end command has been issued or if it is distorted (judgment), and if not, repeat the same process as for the first voice. . If it was,
Since the group has already been determined, segmentation and similarity calculation 2 are performed using only the standard pattern of that group (processing step 2), and the process moves to the phoneme recognition routine. As described above, the processing flow of the device is simple, and the device is characterized in that it can be realized without performing particularly complicated calculation processing. Using the method of this example, 100 adult men and women were
Using the first 10 words out of 212 words, the number of words required to exceed the threshold was determined for each speaker, and the results of evaluating the number of speakers are shown in Table 1.

[Table] In other words, if up to four words are used, up to 98 out of 100 people can correctly judge the group.
One of the remaining two participants required up to nine words, but was able to correctly determine the group. If this group judgment is incorrect, the vowel/nasal recognition rate will be 88.4% → 59.3
%, so it is important to avoid making mistakes by increasing the number of words you learn. I got it wrong 1
This is a case in which a woman is mistaken for a man, but even when this speaker uses the standard pattern for men, the recognition rate for vowels and nasal sounds is 78.5% → 75.5%, which is an extremely small drop in the recognition rate. In other words, the voice of this speaker also conforms to the standard male pattern, so there is no problem even if the speaker is misidentified as male or female. In this way, by using this embodiment, it becomes possible to discriminate between men and women with high accuracy.

[Table] The second table shows the results of evaluating and comparing the average phoneme recognition rate of five vowels and nasal sounds on a frame-by-frame basis for 20 men and women.
Shown in the table. In the case of no distinction between men and women, standard patterns for men and women are prepared separately, and the standard pattern that obtains the maximum similarity without distinguishing between men and women is used as the recognition result, as described in the conventional method. The method according to this embodiment distinguishes between men and women. Each frame recognition rate is shown in %,
The variation in recognition rate is shown in parentheses as standard deviation. Compared to the conventional method, using this embodiment improves the recognition rate and reduces variation. In particular, there was a great effect on improving the recognition rate for women and reducing the variation for men, demonstrating the effectiveness of this example. Effects of the Invention As described above, the present invention divides the voices of multiple speakers into groups in advance, creates a standard pattern for recognition for each group, and uses unknown input voice to find the most suitable pattern for that voice. By providing a function that automatically selects a standard pattern, it is possible to recognize speech using the standard pattern that is most suitable for the user's voice, without putting any burden on the user, and to recognize unspecified speakers. This makes it possible to achieve stable and highly accurate recognition. 2. By limiting the number of standard patterns used to one set, the amount of calculation can be reduced and a speech recognition device with high processing speed can be realized. It has the advantage of

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram showing a conventional speech recognition device, FIG. 2 is a functional block diagram showing a speech recognition device in an embodiment of the present invention, and FIG. 3 is a functional block diagram showing a group of standard patterns in an embodiment of the present invention. FIG. 4 is a flowchart illustrating the automatic selection function. FIG. 4 is a flowchart illustrating an example of the recognition procedure of the speech recognition apparatus of the present invention. 23...Linear predictive analysis processor, 24...Similarity calculation unit, 25...Standard pattern storage unit, 26...
...Segmentation section, 28... Main processor, 29... Selection section, 30... Word dictionary section, 3
2...Standard pattern switching section.

Claims (1)

[Scope of Claims] 1. An acoustic analysis unit that calculates a spectrum or information similar to the spectrum (hereinafter referred to as spectrum information) obtained from input speech for each frame period, and detects speech sections of input speech and segments each phoneme. a segmentation unit that performs segmentation; a standard pattern storage unit that stores in advance a plurality of standard pattern groups classified by speakers with similar characteristics from standard speech signals composed of multiple speakers; and the standard pattern storage unit. a standard pattern switching unit that selects a standard pattern group within the standard pattern storage unit; and a similarity calculation unit that calculates similarity based on a statistical distance measure for each phoneme using the standard pattern group in the standard pattern storage unit and the spectral information. , a processor unit that determines at least a vowel part from the results of the similarity calculation unit and the segmentation unit and selects position information of a frame indicating a stationary part of the phoneme; and a processor unit that corresponds to the position information obtained by the processor unit. Using the similarity calculated by the similarity calculation section,
A cumulative total of the degrees of resemblance of all input voices to the prepared standard pattern group is set for each standard pattern group, and the reliability is calculated by calculating the difference or ratio between the two standard pattern groups with the largest cumulative total. The standard pattern switching unit is controlled to select one of the two standard pattern groups when the reliability exceeds a certain threshold, and to sequentially read out all the standard pattern groups when the reliability does not exceed the threshold. a selection unit; and a word dictionary unit storing a word dictionary to be compared with the phoneme or syllable sequence created by the processor unit using the standard pattern group selected by the standard pattern switching unit or the standard pattern group with the largest cumulative total. A speech recognition device comprising: 2. A patent claim characterized in that the Mahalanobis distance, which is a combination of a covariance matrix common to all target standard pattern groups and an average value set for each phoneme for each standard pattern group, is used as a statistical distance measure. The speech recognition device according to item 1. 3. Claim 1, characterized in that the cumulative degree of resemblance of all input speech to a prepared standard pattern group is determined by the number of standard patterns that have the maximum similarity or the sum of the maximum similarities. Speech recognition device described in section. 4. The speech recognition device according to claim 1, wherein the standard pattern storage section stores a standard pattern group consisting of at least a male voice and a female voice.