JPH04128900A - Voice recognition device - Google Patents

Abstract

PURPOSE:To prevent voice misrecognition due to a noise mixed with an input voice signal by comparing and deciding time-expanded/compressed data between a voice pitch pattern and an input voice analytic pattern. CONSTITUTION:A voice analysis part 3 analyzes the frequency of the input voice signal 2 and outputs the voice analytic pattern 22. A analytic data DP matching part 4 performs DP matching between the pattern 22 and a pattern 23 outputted by a pattern storage part 7 and outputs the time-expanded/ compressed data 25. A voice pitch extraction part 8, on the other hand, finds the voice pitch frequency of the input voice signal 2 at each time by an autocorrelating method to output the voice pitch pattern 28 and performs DP matching between the previously stored pattern 23 and the pitch pattern 26 of the same voice to obtain time-expanded/compressed data 27. A time expansion/compression comparison part 14 converts the data 25 and 27 into vector data, adds differences between all couples of vectors, and outputs the result as a time-expansion/compression similarity signal 15. A decision part 16 makes a decision according to the similarity signal 15.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a speech recognition device.

[Conventional technology]

An example of the configuration of a conventional speech recognition device is shown in FIG.
The input audio signal is frequency-analyzed at t! as shown at 42 in FIG. to tn, and the audio level for each sampling is set to frequency f, H2 to fnH2.
is calculated and output to the analysis data DP matching section 4.

On the other hand, the voice analysis pattern storage section 7 stores in advance a voice analysis pattern 43 of a voice to be recognized, as shown in FIG.

The analysis data DP matching section 4 combines the speech analysis pattern 42 of the input speech signal 2 output from the speech analysis section 3 shown in FIG. is input.

Using the DP matching method (May 10, 1980, published by Kyoritsu Publishing Co., Ltd., Speech Recognition, pages 107-108), which absorbs the time expansion and contraction between these two patterns, all the sounds stored in the speech analysis pattern storage section 7 are compared. death,·
The most similar words were selected as speech recognition words.

[Problem to be solved by the invention]

In the conventional speech recognition device described above, when an input speech signal 2 with a poor signal-to-noise ratio is input, a speech analysis pattern 4 outputted from the speech analysis section 3 as shown in FIG.
2 contains noise 41, and the analysis data DP matching unit 4 matches the noise 41 registered in the audio analysis pattern storage unit 7 with the audio signal pattern 43 that does not include noise.
When P-manching is performed, noise 41 causes erroneous time expansion and contraction as shown in 45, resulting in the recognition of a different voice similar to the desired voice.

An object of the present invention is to compare and determine the time warping data of a voice pitch pattern extracted from an input voice signal and the time warp data of an input voice analysis pattern, thereby detecting errors caused by erroneous time warping due to noise mixed in the input voice signal. It is an object of the present invention to provide a voice storage device that can prevent voice recognition errors and perform highly accurate voice recognition.

[Means to solve the problem]

The speech recognition device of the present invention includes a speech analysis section that receives and frequency-analyzes an input speech signal, a speech analysis pattern storage section that preregisters analysis patterns of speech that require speech recognition, and a speech analysis section that outputs an input speech signal. The voice analysis pattern and the voice analysis pattern stored in the voice analysis pattern storage section are DP
An analysis data DP matching section that generates time-stretched data for each word using a matching method, a speech pitch extraction section that calculates a speech pitch pattern from the input speech signal, and a speech pitch pattern of speech necessary for speech recognition that is registered in advance. A voice pitch pattern storage unit that generates time-stretched data for each word by a DP matching method between the voice pitch pattern output by the voice pitch extraction unit and the voice pitch pattern output by the voice pitch pattern storage unit. A DP matching unit vectorizes the time expansion/contraction data of each word generated by the analysis data DP matching unit and the time expansion/contraction data of each word generated by the audio pitch DP matching unit, and linearly compares the differences between the vectors. It has a comparison section and a determination section that outputs a word with the highest vector open similarity among the time warping similarity signals outputted by the time warping comparison section as a speech recognition word.

〔Example〕

Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

FIGS. 2 and 5 are diagrams showing the time expansion and contraction between the analysis pattern of the input audio signal and the audio analysis pattern registered in advance in the audio analysis pattern storage unit, and FIGS. 5 is a diagram showing the time expansion/contraction between the pitch change of the input audio signal and the audio pitch pattern registered in advance in the audio pitch pattern storage unit, and FIG. Fig. 4 (
B) is a diagram that vectorizes the time warping data in Figure 3, Figure 7 (A) shows the time warp data in Figure 5, and Figure 7 (B)
is a vector diagram of the time warping data in Figure 6, and Figure 8 is a vector diagram of the time warping data in Figure 4.
It is a figure which shows the method of calculating the difference between the vectors of figures (A) and (B).

FIG. 1 shows a microphone 1, a voice analysis section 3 that receives and frequency-analyzes the input voice signal 2 output by the microphone 1, an input voice analysis pattern 22 output by the voice analysis section 3, and a voice that requires voice recognition. A speech analysis pattern storage section 7 in which analysis patterns are registered in advance, and a speech analysis pattern 22 outputted by the speech analysis section 3 and the speech analysis pattern storage section 7.
an analysis data DP matching unit 4 that performs DP matching with the audio analysis pattern 23 output by the input audio signal 2 to generate time expansion/compression data 25;
a voice pitch extraction unit 8 that calculates the voice pitch pattern 26; and a voice pitch pattern storage unit 13 that registers the voice pitch pattern 26 in advance.
, the audio pitch pattern 28 and the audio pitch pattern 26
The audio pitch DP matching unit 10 generates time expansion/compression data 27 by performing DP matching of After the conversion, a time warp comparison unit 14 linearly compares differences between vectors, and a decision is made to output the word with the highest vector viewing similarity among the time warp similarity signals 15 output by the time warp comparison unit 14 as speech recognition words. It consists of a section 16.

FIG. 2 shows a voice analysis pattern 22 of the input voice signal 2 containing noise 21 and a voice analysis pattern 23 similar to the noise 21 of the input voice signal 2 stored in advance in the voice analysis pattern storage unit 7. 25 is a diagram showing the relationship between

FIG. 3 shows the audio pitch pattern 28 of the input audio signal 2 at this time, the audio pitch pattern 26 stored in advance in the audio pitch pattern storage section 13, and the time expansion/contraction data 27.
FIG.

FIG. 4(A) shows the audio analysis pattern 22 of the input audio signal 2.
4(B) shows a vector obtained by plotting the time expansion/compression data 25 of the voice analysis pattern 23 of FIG. 2 registered in advance on a matrix of each sampling time, and FIG. 4(B) shows the voice pitch pattern 28 of FIG. A vector is shown in which the time warped data 27 of the audio analysis pattern 26 is plotted on a matrix of each sampling time.

FIG. 5 shows a speech analysis pattern 32 of the input speech signal 2 containing noise 31 and a speech analysis pattern that is not similar to the noise 31 of the speech analysis pattern of the input speech signal 2 stored in the speech analysis pattern storage section 7 in advance. FIG. 6 shows the time expansion/contraction data 35 with the pattern 33, and FIG. 6 shows the input audio signal 2 at this time.
The audio pitch pattern 38 and the audio pitch pattern 36 stored in advance in the audio pitch pattern storage unit 13 and the time expansion/contraction data 37 are shown.

FIG. 7(A>) shows the vector of the time warping data 35 of FIG. 5, and FIG. 7(B) shows the vector of the time warping data 37 of FIG.

Next, the operation will be explained.

The voice analysis unit 3 analyzes the input voice signal 2 at each time 1. as shown in FIG. Frequency analysis is performed on tn from f, H2 to f
Calculate each frequency component of nH, , and create voice analysis pattern 2.
Outputs 2. The voice analysis pattern storage unit 7 analyzes and registers the voice to be recognized in advance as shown in 23 in FIG. 2.The analysis data DP matching unit 4 stores the voice analysis pattern 22 and the voice analysis pattern storage unit. It performs DP matching with the voice analysis pattern 23 outputted by 7 and outputs time expansion/contraction data 25.

At this time, it cannot be determined from the input audio signal analysis pattern 22 that 21 is noise, so DP matching is performed to match the noise 21 to a pattern 24 similar to the noise 21, which is a characteristic of the audio analysis pattern 23. Time expansion/contraction is performed, and time expansion/contraction data 25 is obtained.

- The voice pitch unit 8 calculates the voice pitch frequency from each time t1 to tn of the input voice signal 2 using an autocorrelation method (May 10, 1980, Kyoritsu Shuppan Co., Ltd., Speech Recognition, 32-34.
42-43) and outputs the audio pitch pattern 28. The voice pitch pattern 28 is generated by comparing the voice analysis pattern 23 stored in advance in the voice pitch pattern storage unit 13 with the voice pitch pattern 26 of the same voice using the DP matching method in the voice pitch DP matching unit 10, and then generating the time expansion/contraction data 27. The 6-hour expansion/contraction comparison unit 14 converts the time expansion/contraction data 25 and 27 into the fourth [J(A) and (
Convert to vector data as shown in B), find differences between vectors sequentially starting from the vector pair at time t1, add up the differences of all vector pairs, and use the result of the addition as a time warping similarity signal 15 (not shown). Output as . This vector conversion is performed using the input voice analysis pattern 22 in FIG. 2 as the vertical axis and the input voice analysis pattern 23 as the horizontal axis.
3 sampling times, t vs. tl, t2 vs. t2
,...Tfu vs. t7, t8 vs. 1. ．． 19.19 vs. tlo, as shown in FIG. 4(A), is plotted on a matrix and the time warp data 25 is vectorized.

In FIG. 3, the input voice pitch pattern 28 is plotted on the vertical axis and the voice pitch pattern 26 is plotted on the horizontal axis, and vectorization is similarly performed as shown in FIG. 4(B). The difference between vectors is determined by defining a vector using orthogonal coordinates and adding the absolute value of the difference on the vertical axis and the absolute value of the difference on the horizontal axis, as shown in FIG.

For example, if the threshold value of the determination unit 16 is "1", the vector difference addition value of the time warped similarity signal 15 is "5" as shown in FIG. 8, which exceeds the threshold value. The time warped similarity signal 15 is discarded.

Next, a case will be described in which a speech analysis pattern that does not include noise is stored in the speech analysis pattern storage section 7.

Similar to FIG. 2, FIG. 5 shows the voice analysis result 32 of the voice analysis section 3.
0 analysis data in which noise 31 is included The DP matching unit 4 uses the voice analysis pattern 33 that does not include noise and is stored in advance in the voice analysis pattern storage unit 7 and the voice output from the voice analysis unit 3. The analysis pattern 32 is input, DP matching is performed, and time expansion/contraction data 35 is output. This time expansion/contraction data 35 is the fifth data because there is no voice analysis pattern similar to the noise 31 in the voice analysis pattern 33 stored in the voice analysis pattern storage section 7.
Speech analysis pattern 32 and speech analysis pattern 3 in Figure (A)
3 are not similar.

On the other hand, the audio pitch extractor 8 extracts each time t of the input audio signal 2.
The voice pitch frequencies from 1 to tn are determined by the autocorrelation method, and a voice pitch pattern 38 is output. The audio pitch DP matching unit 10 compares the audio pitch pattern 38 with the audio pitch pattern 36 of the same audio as the audio analysis pattern 33 using the DP matching method, and obtains time expansion/contraction data 37. In the time warp comparison section 14, the time warp data 35 output from the analysis data DP matching section 4 and the audio pitch D
FIG. 7A shows the time expansion and contraction data 37 outputted by the P matching section 10.

As shown in (B), add the difference between both vectors after death, and
Similarly to the figure, the time warp similarity signal 15 is output to the determination unit 16.

Here, input voice analysis pattern 32 and voice analysis pattern 3
3 is not similar because there is no noise 31, but Fig. 7(A)
, since the vectors are exactly the same as shown in (B),
The inter-vector difference addition value is "0". Since this inter-vector difference addition value is less than the predetermined threshold value "l", the speech analysis pattern 33 in FIG. 5 is output as the speech recognition result. do.

In this way, if the vector-to-vector difference sum value, which is a time-stretched similarity signal, is greater than or equal to a predetermined threshold, it is determined that the degree of similarity is low, and that time-stretched similarity signal is discarded, and the vector difference value is determined to be lower than the threshold. If so, it is determined that the degree of similarity is high, that time-warping similarity signal is held, and the next sum of time-warping similarity signals is compared with the threshold value. This operation is performed for all the words registered in advance, and the last word remaining has the highest degree of similarity and is output as the speech recognition result.

〔Effect of the invention〕

As explained above, when noise is mixed into the input audio signal, the audio pitch extraction unit outputs time-stretched data that has been erroneously DP matched due to the influence of noise in the audio analysis pattern output from the audio analysis unit. By comparing and determining the voice pitch pattern to be used with the voice pitch pattern pre-stored in the voice pitch pattern storage unit using time warp data obtained by DP matching, it is possible to prevent speech misrecognition due to erroneous time warp. It has the effect of improving recognition performance.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIGS. 2 and 5 are diagrams showing the time expansion and contraction between the analysis pattern of the input audio signal and the audio analysis pattern registered in advance in the audio analysis pattern storage unit, and FIGS. 5 is a diagram showing the time warping between the pitch change of the input audio signal and the sound pitch pattern registered in advance in the sound pitch pattern storage unit, and FIG. 4 (A) is a vectorization of the time warping data in FIG. Fig. 4 (
B) is a vector diagram of the time warped data in Figure 3, Figure 7 (A) is a vector diagram of the time warped data in Figure 5, and Figure 7 (B) is a vector diagram of the time warped data in Figure 6. Fig. 8 is Fig. 4(A). FIG. 9 is a block diagram showing an example of a conventional speech recognition device; FIG. 10 is a diagram showing a method for calculating the difference between the vectors in FIG. 7; FIG. FIG. 11 is a diagram showing the time expansion and contraction with the voice analysis pattern registered in the pattern storage section, and FIG. 11 is a diagram showing the voice analysis pattern in the voice analysis pattern storage section. 1...Microphone, 2...Input audio signal, 3...
・Speech analysis section, 4...Analysis data DP matching section,
7... Voice analysis pattern storage section, 8... Voice pitch extraction section, 10... Voice pitch DP matching section, 13
. . . Audio pitch pattern storage section, 14 . . Time expansion/contraction comparison section, 16 . , 23, 33. 43...Speech analysis pattern,
24...Noise-like speech analysis pattern, 2535.4
5...Speech analysis pattern time expansion/contraction data, 26.36
...Audio pitch pattern, 28.38...Audio pitch pattern of input audio signal, 27.37...Time expansion/contraction data of audio pitch pattern.

Claims (1)

[Claims] a voice analysis section that receives and frequency-analyzes an input voice signal; a voice analysis pattern storage section that pre-registers voice analysis patterns that require speech recognition; and a voice analysis pattern output by the voice analysis section and the voice analysis pattern. An analysis data DP matching section that generates time-stretched data of each word using the speech analysis pattern output from the storage section using the DP matching method, a speech pitch extraction section that calculates a speech pitch pattern from the input speech signal, and a speech recognition system. A voice pitch pattern storage section in which voice pitch patterns of necessary voices are registered in advance, and a voice pitch pattern outputted by the voice pitch extraction section and a voice pitch pattern outputted from the voice pitch pattern storage section using a DP matching method. A voice pitch DP matching unit that generates time expansion/contraction data for each word; and a vector that combines the time expansion/contraction data of each word generated by the analysis data DP matching unit and the time expansion/contraction data of each word generated by the voice pitch DP matching unit. a time warp comparison unit that linearly compares differences between vectors, and a determination unit that outputs a word with the highest inter-vector similarity among the time warp similarity signals outputted by the time warp comparison unit as a speech recognition word. A voice recognition device comprising: