JPH067349B2 - Voice recognition system - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method that is advantageously used in a Japanese voice recognition device that recognizes input voice in syllable units.

BACKGROUND ART In the conventional voice recognition method used in a Japanese voice recognition input device for recognizing input voice in syllable units,
When recognizing a syllable that is obtained by decomposing continuous speech into syllables mainly by pattern matching, multiple syllable recognition rates are reduced for each syllable due to variations in vocalization or deformation due to the influence of preceding and following sounds called articulatory coupling. This is prevented by having a standard pattern. In addition, the recognition rate is improved by updating these characteristic standard patterns with new input patterns.

Problems to be Solved by the Invention In the above-mentioned prior art, the standard pattern number of each syllable is not set. Therefore, when learning progresses according to the frequency of the syllable to be input, the number of patterns of each syllable changes. Bias occurs and recognition of syllables with low input frequency becomes extremely poor.

An object of the present invention is to solve the above-mentioned technical problem, to prevent the pattern distribution occupied by syllables of the characteristic standard pattern from being biased depending on the input frequency, and to minimize the fluctuation of the recognition rate. , And to provide a voice recognition method with improved recognition accuracy.

Means for Solving Problems According to the present invention, by calculating the degree of similarity between the input voice and the characteristic standard patterns of a plurality of types of syllables registered in advance, the characteristic standard pattern candidates are ranked in descending order of similarity. In the voice recognition method, which recognizes each unit and corrects the result by collating with a dictionary or external instruction operation such as a keyboard to obtain the final input, the recognition contribution rate depending on the candidate rank and correction Calculated, set the maximum number of feature standard patterns for each syllable, register the minimum feature standard pattern required for recognition in the initial stage, and then input within a range that does not exceed the maximum number of patterns for each syllable. The feature standard pattern that automatically registers the input feature pattern is updated.
The one with the lowest recognition contribution is preferentially updated, and that value is shifted so that the average value of the recognition contributions of the standard feature patterns belonging to the same syllable name becomes the median of the recognition contributions. Then, the speech recognition method is characterized by normalizing the accumulated recognition contribution.

Effects According to the present invention, syllables that are easily affected by the phonemes before and after the articulatory combination and have a statistically poor recognition rate, and syllables that are frequently input (for example, single vowels, voiced plosives, and ra-songs) are Set the maximum number of patterns to a high value, and set the low value for the syllables that apply in the opposite case (frictional sounds such as fricative sounds such as "sha" and plosive sounds such as "ya") efficiently Characteristic standard patterns are added / updated / deleted by automatic registration at the time of input that constitutes a pattern and does not exceed the number given to each syllable.

As a result, the pattern distribution occupied by the syllables of the characteristic standard pattern is not biased by the input frequency, and thus the fluctuation of the recognition rate can be suppressed as small as possible and a stable voice recognition state can be obtained.

Further, according to the present invention, the recognition contribution is obtained by calculating depending on the candidate rank of the characteristic standard pattern and the presence or absence of correction, and when updating, the one with the lowest recognition contribution is preferentially updated. As a result, the feature standard pattern can be optimized easily and quickly.

Further, according to the present invention, the accumulated recognition contributions are normalized so that the values of the recognition contributions are balanced and the memory is effectively used.

Embodiment FIG. 1 is a block diagram showing the configuration of a Japanese voice input device 1 according to an embodiment of the present invention, and FIG. 2 is a flow chart showing the procedure of voice recognition processing in the Japanese voice input device 1. The Japanese voice input device 1 has a function of recognizing continuously uttered voices in syllable units, correcting the recognition result by a dictionary, and then transferring the recognition results to an external device in units of words or the like.

First, the process moves from step n1 to step n2, and a voice signal is input. That is, the voice that is uttered and input is input to the analog input unit 3 via the microphone 2, amplified by the amplifier 4 in the analog input unit 3, and then converted into a digital signal by the analog / digital converter 5. The digital signal is input to the voice analysis unit 6 and the syllable segmentation unit 7.

Next, the process proceeds from step n2 to step n3, and acoustic processing and syllable segmentation processing are performed. That is, the voice analysis unit 6 divides the input voice into frames of about 16 ms, performs spectrum analysis, and transfers the characteristic parameters necessary for syllable segmentation to the syllable segmentation unit 7 at intervals of about 8 ms. The syllable segmentation unit 7 cuts out syllables while temporarily storing various feature parameters from the voice analysis unit 6 in the ring-shaped feature pattern buffer 8 to pattern the features of each syllable and stores them in the feature pattern memory 9. The characteristic pattern buffer 8 is configured to be able to store a plurality of syllables. The processing of the syllable segmentation unit 7 is configured such that start / stop is controlled by an instruction from a central processing unit (hereinafter referred to as CPU) 10.

Next, in step n4, syllable recognition processing is performed.
At n5, recognition result candidates are selected. That is, the syllable recognition unit 11 calculates the pattern distance between the characteristic pattern of each syllable and all the characteristic standard patterns stored in the characteristic standard pattern memory 12 in advance, and outputs candidates in descending order of similarity frequency. The candidates having the same kind of syllable name are integrated and stored in the recognition result memory 13 as a voice recognition result.

Next, in step n6, the error in the syllable recognition result is corrected, and the syllable recognition result determined in step n7 is derived. That is, the correction processing unit 11a in the syllable recognition unit 11 automatically corrects an error in the syllable recognition result using the dictionary stored in the language processing dictionary memory 14. Alternatively, the operator himself may use the keyboard 15 to select a correct candidate from the recognition candidates for the input voice, or may directly correct the error portion. The correct result determined in this way is converted into Kanji and output as a character string.

Next, the learning process is performed in step n8, and the process ends in step n9. That is, in the learning control unit 11 b in the syllable recognition unit 11, learning information updated by analyzing the recognition result by the confirmed character string, specifically, the count value of the recognition contribution counter 16 and the memory 17, are stored in the memory 17. Based on the recent recognition situation, it is determined whether or not to learn the characteristic pattern of each input syllable. For the syllables determined to be learned, the characteristic pattern of the input syllable is added as a characteristic standard pattern, or the worst pattern among the characteristic standard patterns having the same syllable name as the input syllable, that is, the count of the recognition contribution counter 16 is counted. The feature standard pattern is updated by replacing the pattern having the smallest recognition contribution with the lowest recognition contribution with the feature pattern of the syllable, or performing an averaging operation using the feature pattern. In addition, a feature standard pattern having an extremely low recognition contribution may be deleted.

The maximum standard pattern number of each syllable is stored in the maximum syllable pattern number table 18. The CPU 10 controls everything except the voice analysis unit 6.

FIG. 3 is a flow chart showing a more detailed processing procedure of the standard pattern learning processing in step n8. The recognition contribution counters 16 of the empty feature standard pattern areas which are not satisfied by the initial voice registration are all set to the value of Cres in advance. Here, Cres indicates the counter reset value set in the unregistered pattern.

The recognition result of each syllable is analyzed by the syllable Sin determined by the correction processing unit 11a, and the recognition contribution counter 1 is analyzed in step m1.
Update 6 For example, as shown in Table 1, the increase / decrease amount is given according to the candidate rank, and the counter is increased or decreased. That is, when the feature standard pattern having the first candidate rank is correct according to the analysis result, the value of the recognition contribution counter corresponding to the feature standard pattern is incremented by 4, and when the analysis result is incorrect, the recognition contribution is Decrement the degree count value by 4.

For example, Table 2 shows an example in which the correct syllable is "i". That is, the value of the recognition contribution counter Cj is increased / decreased with respect to the pattern that has become the top candidate for recognition of the correct syllable “I”. Here, Cj represents a recognition contribution counter corresponding to the standard pattern Pj.

Next, in step m2, the updated recognition contribution counter C
It is determined whether or not the value of k is less than or equal to the threshold value Cth for deletion, and if so, the characteristic standard pattern Pk corresponding to the recognition contribution counter Ck is deleted in step m3.
At the same time, at step m4, the number n of the current characteristic standard patterns is decreased by 1, and at step m5, the recognition contribution counter Ck of the pattern area is reset to Crec.

Next, in step m2, the updated recognition contribution counter Ck
When the value of is not equal to or less than the threshold value Cth for deleting, the process proceeds to step m6, and it is determined whether or not to learn each input syllable based on the latest recognition situation data from the memory 17. If it is determined to learn, in step m7, the characteristic standard pattern Pm corresponding to the recognition contribution counter Cm showing the minimum value is obtained among the characteristic standard patterns having the same syllable name as the definite syllable Sin. The feature standard pattern Pm has the lowest recognition contribution in the same category.

Next, at step m8, the current pattern number n for the syllable Sin
Is determined to be less than a predetermined maximum number M. n <M
In the case of, it is judged whether or not the count value of the recognition contribution counter Cm showing the minimum value obtained in the step m7 is larger than the value Cres of the recognition contribution counter having an empty pattern area. Sometimes step m10
The input characteristic pattern Pin is added as the characteristic standard pattern P (n + 1). Then, in step m11, the current pattern number n is incremented by 1, and in step m12, the value of the recognition contribution counter Cn corresponding to the characteristic standard pattern P (n + 1) is set in Cset. Here, Cset indicates a counter value set in the added or updated characteristic standard pattern.

If n = M in step m8, the characteristic standard pattern Pm is updated in step m13. Further, in step m9, the recognition contribution counter C showing the minimum value obtained in step m7.
When the value of m is smaller than Cres, step m13
Then, the feature recognition pattern Pm is updated. After the update, the value of the recognition contribution counter Cm is set to Cset in step m14 as in step m12.

Finally, in step m15, the counter values are shifted according to the first equation so that the average value of each recognition contribution counter of the patterns belonging to the same syllable name becomes the median value of the counters to balance the values. This normalizes the cumulative recognition contribution within the same category.

Ci '= Ci + (Ci / L-Ccen) (1) Ci' represents the value of the recognition contribution counter after balancing, Ci represents the value of the recognition contribution counter before balancing, and L is the same syllable. The number of standard patterns of the name
n indicates the median value of the counter variable.

Syllable Sin is “yes”, Cth = 0, Cres = 20, Cs
Tables 3a and 3b show examples of addition of feature standard patterns when et = 50 and the new feature pattern of the input syllable is set to "in". Also, the update example is
It is shown in Tables 4a and 4b. Further deletion examples are
It is shown in Tables 5a and 5b.

(1) Addition Table 3a shows the state before learning, and after learning, the feature standard pattern “I ₅ ” is added as shown in Table 3b.

(2) Update Table 4a shows the state before learning, and after learning 4b
As shown in the second column of the table, the characteristic standard pattern "I ₂ "
Are updated and balancing is performed as shown in the third column.

(3) Deletion Table 5a shows the state before learning, and after learning 5b
As shown in the second column of the table, the characteristic standard pattern "I ₂ "
Are removed and balancing is performed as shown in the third column.

Next, referring to FIG. 4, according to the present invention, as shown in FIG. 4, for example, the total amount of feature standard patterns after learning is , The number of characteristic standard patterns k of j of a certain syllable k
[j] is k [j] = m [j] according to the present invention as shown in FIG.
However, in the conventional case where the maximum number is not set for each syllable, as shown in Fig. 4 (2), Is It is.

In this manner, in the present invention, if the minimum required syllable pattern is initially registered as the characteristic standard pattern, the characteristic standard pattern having a low recognition contribution is within the maximum number of characteristic standard patterns set for each syllable. Add / update /
The characteristic standard pattern can be optimized by performing the basic deletion operation.

FIG. 5 is a graph showing the average recognition rate of syllables for each sentence when 11 kinds of sentences (one sentence consists of 650 syllables on average) are input, and FIG. 6 is the total number of characteristic standard patterns and learned. 6 is a graph showing the transition of the number of input syllables. Here, after registering 590 syllable patterns as feature standard patterns in the initial registration, a maximum number of feature standard patterns of each syllable (1020 in total) was set, and one speaker read the above sentence. It was measured. However, the deletion process is not performed. It can be seen that the feature standard pattern grows and approaches 1020 as the number of input sentences increases. For example, the number of patterns learned when the second sentence is input (the number of syllables that have passed the processing of step m6 in the third step) is added or updated about 300 times, and the number of patterns increased is about 100. . As the number of sentences is input, the average recognition rate becomes higher and the learning effect appears.

EFFECTS OF THE INVENTION According to the present invention, the candidates are ranked in descending order of similarity of the characteristic standard patterns, and the recognition contribution rate depending on the candidate order and the presence or absence of correction is calculated and obtained. Since the lowest one is prioritized, the feature standard pattern can be easily and quickly optimized. This also makes it possible to reduce the memory capacity for storing the characteristic standard pattern.

Further, in the present invention, the maximum number of patterns of the characteristic standard pattern is set for each syllable, thereby increasing the maximum number of patterns for syllables with high input frequency and statistically poor recognition rate, and conversely The maximum number of patterns can be reduced for low patterns, and as a result, the excellent effect that the variation in recognition rate can be reduced with a small storage capacity is achieved.

Further, according to the present invention, by normalizing the accumulated recognition contribution, it becomes possible to effectively utilize the storage capacity.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a Japanese voice input device 1 according to an embodiment of the present invention, FIG. 2 is a flow chart showing the procedure of voice recognition processing in the Japanese voice input device 1, and FIG.
FIG. 4 is a flowchart showing a more detailed processing procedure of the standard pattern learning processing of step n8. FIG. 4 is a diagram comparing the present invention with the prior art. FIG. 5 is 11 kinds of sentences (one sentence consists of 650 syllables on average). ) Is a graph showing the average recognition rate of syllables for each sentence, and FIG. 6 is a graph showing changes in the total number of characteristic standard patterns and the number of learned input syllables. 1 ... Japanese voice input device, 2 ... Microphone, 6 ... Voice analysis unit, 7 ... Syllable segmentation unit, 8 ... Feature pattern buffer, 9 ... Feature pattern, 10 ... CPU, 1
DESCRIPTION OF SYMBOLS 1 ... Syllable recognition part, 11a ... Correction processing part, 11b ... Learning control part, 12 ... Feature standard pattern, 13 ... Recognition result memory, 14 ... Language processing dictionary memory, 15 ... Keyboard,
16 ... Recognition contribution counter, 17 ... Recent recognition status memory, 18 ... Maximum number of syllable patterns table, M ... Maximum number of patterns of each syllable

Claims (1)

[Claims] 1. A method of calculating the similarity between input voices and the feature standard patterns of a plurality of types of syllables registered in advance, ranks the candidate feature standard patterns in descending order of similarity, and recognizes them in syllable units. In a voice recognition method that corrects the word by comparing it with a dictionary or an external instruction operation using a keyboard to obtain the final input, the recognition contribution rate that depends on the candidate rank and the presence or absence of correction is calculated to obtain each syllable. The maximum number of characteristic standard patterns is set in, and the minimum number of characteristic standard patterns required for recognition is initially registered.When inputting, the input characteristic patterns are automatically detected within the range not exceeding the maximum number of patterns of each syllable. Register and update the characteristic standard pattern.
The one with the lowest recognition contribution is preferentially updated, and that value is shifted so that the average value of the recognition contributions of the standard feature patterns belonging to the same syllable name becomes the median of the recognition contributions. Then, a speech recognition method characterized by normalizing the accumulated recognition contribution.