JPH02148100A - voice recognition device - 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 Field of the Invention The present invention relates to a speech recognition device that derives recognition results by pattern matching an input speech pattern and each standard pattern.

2. Description of the Related Art In general, a speech recognition device calculates the degree of similarity between an input speech pattern and each standard pattern stored in a dictionary, and selects the standard pattern with the maximum degree of similarity as the recognition result. Dynamic programming method is used to calculate the similarity between two speech patterns.
Pattern matching (hereinafter referred to as DP matching) that non-linearly expands and contracts the time axes of two patterns using
is used. In particular, word speech recognition devices achieve high recognition rates using this DP matching method. For example, it is described in the following document.

“Dynamic Programming Optimization for Spoken Word Recognition J
(H, 5akoe and S, Chlba, “Dyn
amic pr.

grammlng opt, 11zatlon for
5porken word recognition
”, IEEE t,rans, Acoustlc, 5p
eech, Signal Processing, vo
l, As5P-27pp, 33[1-349, 1979
)) Problems to be Solved by the Invention However, in the above-mentioned speech recognition device, if the section of the input fame is incorrectly detected, erroneous recognition often occurs. Especially in the case of sounds that start with a voiceless consonant (Fukuoka, Fukushima, Fukui, etc.), for example, when you say r F tJ KUOKAJ (Fukuoka), the vowel r U J of rFUJ at the beginning of the word becomes unvoiced. As a result, the energy level at the beginning of a word becomes low, and when detecting a voice section, only the part "rKUOKA" is often detected.

However, when people vocalize, they tend to consciously try to pronounce it clearly in environments with high ambient noise levels, and as a result, the vowel rUJ of [FU-ko] at the beginning of the word becomes voiced. There is also, in this case rFUKUO
The portion KAJ or r U K UOKAJ will be detected as a voice section. In this case, when there is a possibility that the beginning of a word is missing depending on the situation during pronunciation, if the speech interval detection results of the standard pattern registered in advance in the dictionary and the input speech pattern are different (for example, one is r F U K
(UOKAJ and the other is detected as rKUOKAJ), the conventional DP matching method cannot cope with this problem and causes misrecognition.

In view of the above-mentioned problems, the present invention provides a speech recognition device that can accurately recognize even when input speech with a possibility of missing the beginning of a word (especially speech that is likely to be devoiced at the beginning of a word) is input. be.

Means for Solving the Problems The speech recognition device of the present invention includes a fame analysis section that outputs a time series of feature vectors including an energy value time series from input speech, and a reputation analysis section that outputs a time series of feature vectors including an energy value time series from an input speech, and a reputation analysis section that outputs a time series of feature vectors including an energy value time series from an input speech. A silent section detecting section detects a silent section and stores the starting point of the subsequent energy, and a starting point is the starting point obtained as a result of voice section detection and the first silent section point of the incoming cover turn and standard pattern, respectively. The present invention is characterized by comprising a pattern matching section that performs pattern matching calculations on candidate points.

Effect of the Invention With the above-described configuration, the present invention takes into account the omission of the beginning of a word, and uses the start point of the first silent section point on the time series of the speech pattern as well as the start point of the standard pattern and input speech pattern obtained by speech section detection. By starting pattern matching calculations as candidates and finding the similarity between the two,
Accurate recognition is possible even when recognizing speech that may be missing the beginning of a word.

Embodiment Hereinafter, a speech recognition device according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a speech analysis unit which converts an input speech signal into a time series of feature vectors including a time series of energy values and extracts speech sections. Reference numeral 2 denotes a silent section detecting section which detects a silent section from the audio energy value time series and stores the frame position of the subsequent energy rise as a silent section point. 3 is a switch for switching between registration and recognition operations; 4 is an input cover turn memory that stores the time series of feature vectors of input speech and their silent interval points;
Reference numeral 5 denotes a standard pattern memory that stores the time series of feature vectors of each speech to be recognized and the respective silent interval points. 6
is a pattern matching unit that performs pattern matching using the starting point obtained by voice section detection and the first silent section point as starting point candidate points for pattern matching calculation.

7 is an initial setting unit that sets the distance between vectors of points on the main axis of pattern matching other than the start end candidate point to infinity.

Reference numeral 8 denotes a cumulative distance calculation control unit that sequentially calculates the cumulative distance from the first frame of the input audio pattern using an asymmetric matching path that does not depend on the length of the standard pattern, and the pattern matching meter p corresponds to the silent section of the input audio pattern. When the frame corresponding to the point is reached, cumulative distance calculation is performed in parallel starting from the silent section point. 9 is a result output unit which outputs the standard pattern with the maximum degree of similarity as a recognition candidate speech. FIG. 2 is an explanatory diagram of the apparatus shown in FIG. 1.

Next, the operation of the above speech recognition device will be explained.

First, the audio signal input from a microphone, etc. is converted from analog to digital in the audio analysis unit 2, and converted into a time series of voice feature vectors (for example, 10th linear prediction coefficient) and an energy value time series, and the uttered voice The part is detected as an interval. Next, the silent section detecting section 3 detects, as a silent section, a section in which the energy value is below a predetermined threshold for a certain period of time from the energy value time series obtained by the speech analyzing section 2, for example. The frame position of the rise of energy following the silent section is stored as the silent section point.

When registering speech to be recognized, set the switch 3 to the T side and use the time series of feature vectors obtained by the speech analysis section 1 and the silent section points obtained by the silent section detection section 2 for each speech to be recognized. It is stored in the standard pattern memory 5.

Next, a description will be given of the case where the switch 3 is set to the R side, that is, the recognition processing operation. The input speech is subjected to the above-described processing in the speech analysis section 1 and the silent section detection section 2, and the time series of its feature vectors and the frame positions of the silent section points are stored in the input cover pattern memory 4. Next, the pattern matching section 6 sets the start end and the first silent section point as start end candidate points, and executes pattern matching processing based on dynamic programming. Below is the pattern matching part 6
The operation will be explained in detail.

Here, we consider the following equation as a recurrence equation for pattern matching.

g(1, I): d(1, I) −(
1) g(i,j)=min(g(1-Lj),g(
1-Lj-1), g(1-1,j-2))+d(j,
j) -----(2) Here, g(i
, j) and d(i, j) are the i of the input speech pattern.
frame, the cumulative distance in j frames of the standard pattern, and the distance between feature vectors. Note that the recurrence formula may be other than -1 as long as it realizes asymmetric route restriction.

Furthermore, for simplicity, it is assumed here that restrictions on matching paths by matching windows are not considered.

- As an example, consider the case where pattern matching is performed between an input speech pattern and a standard pattern as shown in FIG. In FIG. 2, the envelope for the speech pattern shows the change in energy over time. Here, it is assumed that the frame position of the first silent section point of the input audio pattern is IQ+, and the frame position of the first silent section point of the standard pattern is jo.

By the way, especially when the speech to be recognized starts with a voiceless consonant, there is a possibility that the initial section of the word will be detected incorrectly in both the standard pattern and the input speech pattern.

In this way, when there is a possibility that the beginning of a word is missing in both patterns, the starting point of the pattern matching path is (1+1), (tQl+IL) as shown in Figure 2.
Three starting point candidate points of (1, job) are considered, and by selecting the optimal path from the three matching paths of ■■■ starting from these three points, it is possible to realize pattern matching that can cope with the missing beginning of a word.

Therefore, first, the following equation is set in the initial setting section 7.

d(1,j)=■ However, j≠L j≠jo+ d(i, 1)=■ However, i≠1.1-I-1Q.

This setting is equivalent to adding a restriction so that the matching path does not start from any point on the straight lines i=1 and j=1 in FIG. 2 other than the three starting point candidate points.

Next, the cumulative distance calculation control unit 8 starts calculating the cumulative distance according to recurrence formula (2). First, with i=1, the cumulative distance is calculated in the vertical direction from the first frame of the standard pattern, that is, j=1. When this is completed, the same process is continued while sequentially incrementing i.

After these processes progress and reach j”!o+,
The cumulative distance calculation starting from the point (i 0++ 1 ) is performed in parallel with the previous calculation of g(i, j).

That is,! :! After at, at the same time as calculating g(i, j),
Calculation of g'(Lj) according to the following recurrence formula (11) is performed while incrementing i.

When these processes reach point (I, J), two cumulative pressures #g(I, J) and g'(1, J) have been determined. Here min (g(1,J)/l, g'(1,J)/(1
−1o++l)1, and the obtained value becomes the degree of similarity between the input speech pattern and the standard pattern. This degree of similarity is the distance between patterns when the optimal route is selected from among the three types of matching routes started from the three starting end candidate points in FIG.

After all the degrees of similarity between each standard pattern stored in the standard pattern memory 7 and the input voice pattern stored in the input voice pattern memory 4 are calculated, the result output unit 9 recognizes the standard pattern with the highest degree of similarity. It is determined as a candidate voice and output to the outside.

As described above, according to this embodiment, the silent interval detection unit 2 that detects a silent interval of speech and the distance between vectors at points on the principal axis - other than the plurality of start end candidate points that take into account word-initial omissions are set to infinity. The initial setting section 7 sets the cumulative distance calculation to be newly started from the first silent section point of the input audio pattern.
By being equipped with the cumulative distance calculation control unit 8 that controls the cumulative distance calculation to be executed in parallel with the cumulative distance calculation starting from the intersection of the starting edge of the input speech pattern and the standard pattern, it is possible to determine the beginning of a word in either the standard pattern or the input speech pattern. Even if a missing word occurs, calculations can be performed that consider all possibilities with a single pattern matching operation, and even speech with unstable beginnings of words can be recognized accurately without increasing the amount of calculation. becomes possible.

As described above, according to the speech recognition device of the present invention, the start point of the standard pattern obtained by speech interval detection and the starting point of the input speech pattern as well as the starting point of the speech pattern are By starting pattern matching calculations using the first silent interval point as a starting point candidate and finding the degree of similarity between the two, it is possible to accurately recognize speech that may be missing the beginning of a word. .

Specifically, as a pattern matching part, there is an initial setting part that restricts the matching path from starting from a point other than the starting point candidate point of multiple pattern matching that takes into account missing word beginnings, and the intersection of the input cover pattern and the starting end of the standard pattern. By being equipped with a cumulative distance calculation control unit that concurrently performs cumulative distance calculations starting from , and cumulative distance calculations starting from the silent section point of the input kataan, it is possible to eliminate the problem when there is an unstable sound at the beginning of a word in the speech to be recognized. Regardless of whether the beginning of a word is missing in the input speech pattern or the standard pattern, all pattern matching that takes into account the missing beginning of a word can be achieved in one calculation, improving recognition performance with almost no increase in the amount of calculation. can be done.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention, and FIG. 2 is an explanatory diagram of the operation of the same device. 1. Fist and fist audio analysis section, 2. Fighting fist silent section detection section, 3.
and switch, 4 middle...Input cover turn memory, 5 fight...Standard hat turn memory, 6◆・Fist pattern matching section, 7
...Initial setting section, 8/9 fist cumulative distance calculation control section, 9.
-Dispute result output section.

Claims (2)

[Claims] (1) A speech analysis unit that outputs a time series of feature vectors including an energy value time series from the input speech signal and detects the uttered speech interval; a silent section detecting section that detects a silent section in the pattern and storing the rising position of the subsequent energy as a silent section point, and the speech analyzing section for each of the standard pattern and input speech pattern that are stored in advance as speech to be recognized. The beginning of the detected voice section,
The similarity is calculated between the standard pattern and the input speech pattern by using the first silent section point in the speech pattern time series detected by the silent section detection unit as the starting end candidate point of the speech pattern, and the similarity is calculated as the maximum similarity. A speech recognition device comprising: a pattern matching section that outputs a standard pattern as a recognition candidate speech. (2) The pattern matching unit calculates only the distance between the feature vectors between the starting end candidate points of the standard pattern and the input audio pattern, and calculates the distance between the feature vectors of the starting end candidate points of the standard pattern and the input audio pattern, and calculates the distance between the feature vectors of the starting end candidate points of the standard pattern and the input audio pattern and the input audio pattern. An initial setting section that sets all the vector distances between points other than the first frame of the input audio pattern and the starting end candidate point of the standard pattern to very large values in advance, and calculates cumulative distances sequentially from the first frame of the input audio pattern, It is characterized by comprising a cumulative distance calculation control unit that simultaneously calculates the cumulative distance starting from the first frame and simultaneously calculates the cumulative distance starting from the first silent section point of the input audio pattern when the section point is reached. The speech recognition device according to claim 1.