JPS5868098A - 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] The present invention is particularly characterized by learning pattern creation means,
This invention relates to a speech recognition device that uses dynamic programming as a means of pattern matching.

In a speech recognition device, an input speech is analyzed to extract a feature vector, the feature vector is obtained in a specified time unit, and the feature vector/zetern is learned and memorized. The input voice is identified by turn matching.

In a speech recognition device that uses conventional dynamic programming as a means of pattern matching, a feature vector representing each unit time extracted by a means for analyzing input speech is compared with the feature vectors before and after it in time. Learning patterns have been formed regardless of similarity.

If multiple 101 learnings are performed, the time-corresponding feature vectors among them are averaged, or separate learning patterns are formed for the same training speech, and during recognition, it is treated as if it were a different speech. I have been performing independent recognition operations from time to time, as if it were learning or turning.

Furthermore, the input speech during the recognition process was used primarily for comparison with the nine turns of learning, and did not affect the learning pattern.

The sounds a person makes each time they are uttered become stronger. Therefore, in many cases, the human-generated voice in the recognition process is different from the human-generated voice in the learning process, but with the learning and Zetaan formation means using the method described above, the change will be noticeable in the pattern matching results. It was not possible to improve the recognition rate beyond a certain limit.

Naive, when learning multiple times, it is difficult to reflect the results of multiple learnings in one learning pattern, and the method of forming separate learning patterns depending on the number of learning times reduces the memory W m There is a problem that the p1 response becomes large and the p1 response takes a long time due to the large number of operations during the five-e-knowledge process.

As mentioned above, human vocalizations are not uniform and change from moment to moment. Therefore, when using a learning pattern that was learned more than one month ago, for example, it often causes problems in practical use.For this reason, in conventional methods, although the recognition device is used continuously, learning is performed frequently. must be carried out.

This invention was made in view of the above points, and it is possible to absorb fluctuations in input speech caused by changes in vocalization, effectively obtaining stable recognition results, and, for example, in limited verification. It is possible to perform reliable recognition calculations within a certain range, speed up recognition operations, and effectively measure the speed of recognition operations.In particular, learning e-turns corresponding to the speaker's current utterances are created, and the recognition rate decreases over time. The present invention seeks to provide a 7:1-voice recognition device that reliably prevents the use of voice recognition.

That is, in the present invention, in the speech learning process, the unit time (for example, 1
For the feature vectors representing the speech of 0 msec), the total similarity between the temporally continuous feature vectors is A,
As a result, if the similarity between adjacent feature vectors is above a certain value, it is assumed that the same voice has been input continuously, and the averaging operation is performed on the feature vectors, and this feature vector f: + the equalized feature vector Learning patterns are created by replacing or averaging feature vectors and the number (duration) of averaged feature vectors by replacing them with VC, and dynamic programming is used as a means of gloss turn matching. .

In other words, by taking five steps to create such a learning pattern, it becomes easy to reflect the effects of multiple learnings, and to re-learn voices for which the learning pattern has already been learned. In this case, pattern matching is performed between the input speech and the learning pattern using dynamic programming, and the feature vectors of the input speech and the learning pattern that are matched by the time deviation relationship obtained in the process are A weighted average is taken based on the number of learning times, and the resulting time-series feature vector sequence is averaged in the time direction. Therefore, by performing multiple learning and creating learning patterns and turns, a more average learning pattern is created, resulting in stable recognition #i! i
The fruit is good? In addition, by training a large number of different people, the speech recognition of unspecified speakers can also be made to be ``speech-like''.

Furthermore, in the voice recognition process, if it is determined that the human voice is the same as the voice that has already been learned, the voice input for recognition will be treated as VC in the same way as the human voice for multiple learning described above. - It is designed to make the user learn, and regularly performs voice recognition at 11
Now J'Lula learns and learns a part of the pattern each time it is used, and always uses audio waves II #! +ic M good learning 9
It's a good idea to make it last for one turn.

An embodiment of the present invention will now be described with reference to the drawings. m
1 i'n] is the meaning of (I'・"),
Reference numeral 11 denotes a microphone that detects the input sound (1... j is increased by 1 to 1 by 1 and supplied to the low-pass filter 13. Then, the input line j
The audio wave 1 of K. The value 51 is detected by the CPU 15. This CPU 15 is supplied with an interrupt signal for sampling from the timer 16, and this interrupt signal causes the A//b converter 14 to be activated by the A interrupt.
The digital value VC-converted audio waveform data obtained from the digital value VC conversion is performed. And CPU 15
Then, a preferred learning pattern is created and stored for this sampling data, and a speech recognition operation is performed by pattern matching the learning pattern and a pattern based on input speech, and the output device 17 is driven and controlled according to the recognition result. be.

In a speech recognition device configured in this way, first a learning pattern for speech recognition is created.
Further, a recognition operation is quickly performed on this learning pattern, and the learning process and recognition process will be explained with reference to the flowcharts shown in FIGS. 2 and 3, respectively.

1. When the learning process is started, the process proceeds to step 100, in which the signal input from the microphone 11 is input to the amplifier 12 by an interrupt signal from the timer 16.
The audio waveform converted into a digital value via the low-A space filter 13 and the A/T converter 14 is sampled. When a certain number (for example, 256 pieces of data, hereinafter referred to as a frame) of data is sampled, the process proceeds to step 101, where the data is analyzed. As a result of the analysis of this data, a small number of parameters (assumed to be F) representing the sampling data are extracted, and are called a feature vector and expressed using the following equation.

■""(" + v"' + 1)t, f+ "'
＋? Jt, F1t t, i t2
'Here, ■, is the feature vector of the ti-th frame, and υ, f is its f-th feature vector component 1.

Also, even during execution of data analysis, an interrupt from the interrupt signal from the timer 16 causes step 1 to be interrupted.
00, data sampling of the next frame is executed.

When the data analysis is completed, the process proceeds to step 102, and it is determined from the feature vector obtained as a result of the analysis whether the input has been completed. The end of input is determined based on the following conditions a and b, for example.

(,) Calculate the power of the input voice from the % signature vector9- and determine whether the input can be determined to be valid input or whether it has been input a certain number of times or more.

(b) Whether or not there is a certain number of consecutive inputs of small power that can be determined to be that there is no input audio.

If this judgment result is “NO”, step 100゜10
Proceed to step 1 and perform data sampling and data analysis again. If "YES", the process proceeds to step 103, where it is determined whether the input voice has already been learned. If the result of this determination is lN0J, the process proceeds to step 107, where averaging processing is performed. If the answer is "YES", the process proceeds to step 104, where a pattern matching process is performed between the feature vector sequence of the input voice and the learning pattern.

Pattern matching is performed using dynamic programming (dynamic zologramming, hereinafter referred to as DP), which matches the starting and ending ends of the input pattern and the learning pattern.

As a result, by a defined distance measure (described below),
The distance between voice patterns is determined. (The smaller L is, the more similar it is.) Also, the input cut + turn 10
- Time warping function that represents the correspondence between the feature vector and the feature vector of the learning pattern
are determined at the same time. An example of a time deflection function is shown in FIG. In the figure, T and Tp are the number of feature vectors of the input pattern and the learning pattern, respectively.

Next, the process proceeds to step 105, where it is determined whether the input voice and the learning voice are the same. This is a predetermined threshold value for the distance between the voices. 1, if [7 and L<L, 11, it is determined that they are the same, and anything other than this Jt is determined to be not the same. If it is determined that they are not the same, it is determined that the input voice is different from the learning voice and is an invalid input, and the learning process is terminated. If it is determined that they are the same, step 10
Proceed to step 6 to perform multiple learning processing.

The process of this multiple learning process will be explained in detail according to the flowchart shown in FIG. That is, when entering the multiple learning processing process, the process proceeds to step 116, where the index t indicating the feature vector number of the input pattern is set to 11'', and step 1
26 Vc, and using the time deflection function W obtained in step 104, set the inertia vector number j of the learning pattern corresponding to the summation vector number t of the input kataan, and similarly, - times ago. Set the feature vector number j of the learning pattern of the operation.

Next, the process proceeds to Stella 7'J166, where a weighted average of the feature vector V of the input pattern and the feature vector Vj of the learning pattern is performed. Assuming that the learning number is N times,
For each feature vector element. Here, V a , 1 is the t-th element of the ji-th feature vector of the learning pattern, υ).

is the first element of the t-th feature vector of the input Kanoya turn, and υj9. is the t-th element of the averaged feature vector of the j-th order. Furthermore, the learning pattern consists of a feature vector and the number of consecutive feature vectors as described below, and in the multiple learning process, the learning pattern is expanded and calculated. Further, the upper limit of N is set as 17'', and when N is less than 6, 11'' is added to N.

Next, the process proceeds to step 176VC, and it is determined whether or not averaging has been completed for all feature vectors of the input pattern. “If YESj, end the multiple learning process.
If NOJ, proceed to step 186, add 11'' to t, proceed to step 136, and similarly to step 126, time-deviate the number of the feature vector of the learning pattern corresponding to the feature vector V of input turn 1. Find it from the function W and set it to j. Next, the process proceeds to step 146, and it is determined whether the strength of the feature vector of the learning pattern obtained in step 136 is equal to the number of the feature vector of the learning pattern that was averaged last time. [In the case of YESJ, the feature vectors l and (i-1) are determined to be almost the same by DP, so again? -'i determines that it is not necessary to incorporate the same feature vector into the learning pattern, and proceeds to step 13-176. 1 NO'', proceed to step 156yc, set the value of j in jo, and proceed to Stella f166.

The following process is similar to the process described above.

When such multiple learning process is completed, the process proceeds to the averaging process of step 107 in FIG. 2.

The process of this averaging process will be explained in detail according to the flowchart shown in FIG. The index indicating the column number is also set to "2", and the index indicating how many consecutive feature vectors determined to be the same voice is 1. Feature vectors that are registered as different feature vectors as a result of averaging. Set each number index to "1". Next, step 17
Proceed to step 2, and set the feature vector VR to be compared as v, (
If it is the first time, set it to ■, and so on). Next, proceeding to step 173, the distance i between the feature vectors VR and V
Perform the I calculation and set the value to R. The distance calculation method depends on the data analysis method, but for example, when frequency analysis is performed, the Euclidean distance of the logarithm is calculated, and when Quezonutrum analysis is performed, the distance is calculated using the Euclidean distance. For the Euclidean distance of the Cepnutrum coefficient, the likelihood ratio or coiIh scale is used when linear predictive analysis is performed. When performing a 1-core (PARCOR/) analysis, the following Euclidean distance of 9-core coefficients is used.

Next, the process proceeds to step 174, where a predetermined threshold value is calculated for the distance Dt obtained by the above calculation method. Determine whether the value is greater than or not. In the case of NOj, it is determined that ■8 and v2 are feature vectors representing the same voice, and the process proceeds to step 175, where the weighted average of ■, and ■ is calculated.

Averaging is performed for all components of the feature vector.

(f-1 to F) Next, the process proceeds to step 176VrL, and it is determined whether the comparison has been performed up to the last feature vector.

If I'YESJ, go to step 184. l'
- In the case of NOJ, "1" is added to each of t and k and the process returns to step 173.

In the above step 174, if i'YEJ, the feature vector vR and ■□ are the same sound 7! - is determined not to be representative, and the process proceeds to step 178.
Vya is set as the feature vector (7) of 1L block, and the value of k is set in the number of fc feature vectors that are determined to be equal to the feature vector (I).

Next, the process proceeds to step 179, in which it is determined whether or not the comparison has been performed using the last feature vector 1, as in step 176.
If "YES", the process advances to step 184. If "NO", proceed to step 180, reset I(to "1", and set it to 1 and t in step 180).
＋I get it. Next, proceed to step 182vC, update the feature vector ■□ to be compared to a new ■, and Stella:
At step 7'183, "1" is added to t and the process goes to step 173.

"In steps 176 and 179 of the first paragraph, when it is determined that all feature vectors have been compared, step 184
Don't worry, set the i-th % characteristic vector of the learning pattern to 7, set all -CvR, set the value of k to M, and complete the averaging process. As a result of this averaging process, the learning pattern is converted into feature vectors and the duration of the voice represented by these (the number of feature vectors), and when these are converted into clothes, the result is as shown in Figure 7. .

When the averaging process is completed, step 108 in FIG.
The learning pattern shown in FIG. 7 is registered as the learning pattern of the designated voice, and the learning process is completed.

Next, the recognition process will be explained according to the flowchart shown in FIG. Data sampling, data analysis, and determination of completion of input for Stella 7'' 100, 101° 102 are the same as the learning process shown in FIG. , pattern matching is performed using DP between the input cover turn of the input voice and all the learning patterns. Through this process, the learning pattern with the smallest distance from the input cover turn is selected. Here, the i-th Let the distance between the learning pattern and the input cover turn be Li.Next, proceed to step 111, and determine whether there is a learning pattern that can be determined to be the same as the input cover turn.The determination is made as follows. In other words, LL1 to L, J
(I is the number of learning patterns), the smallest value is LB (
Let the learning pattern having that value be B (the ffth learning iR pattern), and the next smallest value be L, /.

and a predetermined threshold. 2゜Lo6
For this, if both conditions (L.2(1) LB-LB'<L.3(2) are satisfied, the B-th learning pattern is determined to be a learning pattern of the same voice as the input voice. If the condition (1) above is not met, it is determined that a voice different from any of the learning voices has been input, and if the condition (2) is not satisfied, the input voice is the B-th learning pattern. No. 18- It is determined that there is a high possibility that it is a learned voice, but it cannot be clearly distinguished from other learned patterns.

If it is not determined that there is a learning pattern corresponding to the input cover turn, the process proceeds to step 116, and outputs an output indicating that there was effective manpower but there was no learning pattern corresponding to it, or 1, that it was indistinguishable. to the output device and complete the recognition process. If it is determined that there is a corresponding learning pattern, the process proceeds to step 112 and a predetermined threshold 4ttiL is determined. For 4, LB (
L, 4 is determined, and it is determined whether or not the human-generated voice should be learned for recognition.

however L12<L. 2 Suppose that

Here, in the case of 1NOJ, it is determined that the input voice is equal to the scholar voice of the Bth learning pattern, but the learning! Since there is a difference between the turn and "below a certain limit", it is determined that learning is not to be performed, and the process proceeds to step 116, where the input sound %j is
(The 19th learning: The recognition process is finished by outputting to the output interpolation that it is the 87th pattern of learning speech. [-YESJ's 1
The input voice is determined to be the same as the learning voice of the B-th learning pattern and has been uttered with sufficient care;
Assuming that the distance LB is a normal vocalization variation, step 1
The process proceeds to step 13, where multiple learning processing is performed. The processing process is the same as the learning process, so =! 't, F3A is omitted. When operating the device continuously through this process,
1/8 of the learning pattern will be changed.

Next, the process proceeds to step 114, where a feature vector averaging process is performed. This process is also the same as the learning process, so the explanation will be omitted. Next, proceed to step 115 to update and register the B-th learning t'i pattern, and step 11
At 6'J1, the recognition process is completed by outputting to the output device that the input speech is the B-11j learning pattern speech.

Note that the averaging of feature vectors is performed using PA as a feature vector.
Combining using RCOR coefficients is particularly effective due to its linear interpolability, but it is also possible to use
Similar effects can be obtained by using frequency components.

In the above embodiment, the input voice is supplied to the low-pass filter 13 to sample the voice waveform, but as shown in FIG. A plurality of bandpass filters 1B-, 1°18-2. having different pass frequency bands. ...1 B-
In addition to supplying FK, each gundo pass filter 1
The output signals from each of 8-1 to 18-F are detected by a detector 19.
-1 to 19-F may be detected to sample each station wave number component of the input audio. In this case, sampling data of each frequency component obtained from each of the detectors 19-1 to 19-F is sequentially read by a multiplexer 20 controlled by an interrupt command from the CPU 15, and
The D converter 14 intersects the digital value and connects it to C1)U15, and by using each sampling value as a feature vector, the same effect as in the previous embodiment can be obtained.

In the first embodiment, a means for comparing temporally adjacent feature vectors is shown, but it is also possible to reduce the ratio of adjacent feature vectors by one. Furthermore, in the example, the averaging process was performed after the multi-scientist process, but is this true? ', 1 averaging process may be performed first, and then multiple learning may be performed by performing the averaging operation on the corresponding feature vectors.

As described above, according to the 11-voice recognition system according to the present invention, the degree of similarity of temporally adjacent feature vectors is set to =-, and the feature vectors represent the same origin. If it is determined, the averaged feature vectors are averaged, and the averaged jL is used as a substitute for the same voice to create a learning pattern.
It is possible to absorb fluctuations and fluctuations in the speaker's utterance at the time of 1ljk, and it is possible to avoid giving too much S+ value to the fluctuations, so that a stable recognition rate can be obtained. If you want to learn multiple times, first use the input voice and the song with 9 turns to learn Dynamic H111! -C pattern matching is performed using the II method, and the deicing time is = 22. Multiple learning is performed by associating the input speech with the feature vector of the learning pattern and performing an averaging operation using the polarization function. can be performed accurately and can absorb fluctuations in each utterance.In addition, if it is determined during the recognition process that there is a learning pattern that corresponds to the input voice, the system can usually be used frequently by learning the input voice. Learning that is regularly required when using the device □ Operation can be omitted, and part of the learning pattern can be updated sequentially (by constantly updating the learning pattern using the current older voice). It is possible to maintain a turn and maintain a stable high recognition and i!2i rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram explaining a speech recognition device according to an embodiment of the present invention, FIG. 2 is a flowchart explaining the learning process of the device, FIG. 3 is a flowchart of the recognition process, and FIG. 4 is a time diagram. A diagram showing an example of a polarization function, FIG. 5 is a flowchart explaining the multiple learning process, and FIG. 6 is a flowchart explaining the half-equalization process.
FIG. 7 is a block diagram of an averaged learning pattern, and FIG. 8 is a block diagram illustrating another embodiment of the present invention. 1)...Microphone, 13...Pass filter, 1
4...A/I) converter, 15...CPU, 7(li
...Timer, 17...Output device fft, 1B-1~1
8-F...Band pass filter, 19-1 to 19-F
...Detector, 20...Multiplexer. Applicant's representative Patent attorney Takehiko Suzue Figure 1 Figure 2 Figure 3 H

Claims (3)

[Claims] (1) A means for analyzing the sound input to the microphone and extracting a feature vector representative of the input sound for each sampling unit time, and a temporally continuous feature in the sampling time series feature vector sequence extracted by this means. comprising means for calculating the similarity of vectors, and means for determining that the similarity obtained by this means is greater than or equal to a specified value and averaging the feature vectors, and averaging the feature vectors. A speech recognition device characterized in that a learning no or turn is created based on the duration of the speech represented by the fc% feature vector or the averaged feature vector. (2) A means for determining whether the input voice is the same voice in a recognition operation between the input voice and the learning pattern that has already been created, and a means for performing an averaging operation between the corresponding feature vectors of the input voice and the learning pattern at the time of this identity determination. 2. The apparatus according to claim 1, wherein the averaging means is performed using the averaged feature vector sequence to create a learning pattern by multiple learning. (3) A means for determining whether the input voice during the recognition process is the same as the learning/turn, and a means for performing the above-mentioned multiple learning on the input voice during the same determination, and the learning pattern is updated sequentially. An apparatus according to claim 2.