JPH01321498A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To suppress misrecognition and to improve the rejection performance at the time of the input of a word which is not to be recognized by detecting a voiceless section of a standard pattern previously and carrying out pattern matching while storing a previous passing point in a back pointer. CONSTITUTION:A speech analysis part 2 extracts a time series of feature vector from an input speech signal and a voiceless section detection part 3 detects voiceless parts from a time series of energy values of a voice and stores following frame positions where energy rises. Then a similarity calculation part 4 matches the input pattern with patterns whose paths are limited at intersections of the voiceless sections among the standard patterns. Further, a back pointer storage part 6 stores intersections of voiceless section points which are passed before respective points are reached at the time of cumulative distance calculation and a recognition and decision part 7 regards the standard pattern which indicates the shortest distance among distance values obtained by the similarity calculation part 4 as a recognition candidate speech. Consequently, misrecognition among similar words is suppressed and even when a word which is not to be recognized is inputted, it can be rejected as much as possible.

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, ``Dynamic Programming Optimization for Spoken Word Recognition J (H,5akoe and S, Chlba
＋"Dynamlc program mng o
ptlmlzatlon for 5porken
IEEE
trans, Acoustlc, 5peech, Si
gnal Processing, yol, As5P
-27pp, 33G-349゜Problems to be Solved by the Invention However, since the above-mentioned speech recognition device performs pattern matching based only on the shape of the spectrum, even in DP matching between different speech patterns, extreme time axis Due to non-linear expansion and contraction, the distance between the two may become small, causing misrecognition. for example, "
Looking at the energy series of "Osaka" and "Oita", the former has three energy peaks and the latter has two energy peaks, so they are clearly two different patterns, but they are phonetically similar, so DP matching The distance may become small and misrecognition may occur.

In addition, in DP matching between similar speech patterns,
Since DP matching is performed over the entire speech pattern, the difference between the two is obscured, and as a result, the distance between the patterns becomes small and misrecognition is likely to occur.

Furthermore, even when a word that is not to be recognized is input, it is matched to one of the standard patterns, and there is a limit to the ability to paste the word that is not to be recognized.

In view of the above-mentioned problems, the present invention is a speech recognition device that is capable of suppressing misrecognition between similar speech patterns and misrecognition caused by extreme P matching as much as possible, and further improving query performance when a word not to be recognized is input. It provides:

Means for Solving the Problems The speech recognition device of the present invention includes a speech analysis section that outputs a time series of feature vectors including energy sequences from input speech, and detects silent intervals from the energy series output from the speech analysis section. A pattern matching is calculated on all possible paths to limit the matching path at the intersection of the silence section points of the input pattern obtained by the silence section detection section and the standard pattern, and the silence section points of the input pattern obtained by the silence section detection section are calculated. The present invention is characterized by comprising a similarity calculation unit that stores as a back pointer which intersection of silent interval points has been passed before when calculating the cumulative distance.

Effects of the present invention With the above-described configuration, the silent section of the standard pattern is detected in advance, and the pattern matching is executed while storing the previous passing point in the back pointer. By being able to calculate distances that include how to match all silent interval points in consideration of point detection errors, it is possible to eliminate misrecognition between similar patterns and errors caused by extreme matching without increasing the processing time required for pattern matching. Recognition can be suppressed as much as possible, and furthermore, words that are not recognized can be rejected as much as possible.

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, 1 is a microphone for inputting audio signals (it may also be a telephone handset, etc.).
2 is a speech analysis unit that extracts a time series of feature vectors from the input speech signal. Reference numeral 3 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. 4 is a similarity calculation unit that performs pattern matching with a path limited at the intersection of the silent section points of the input pattern and the standard pattern. 5 is a storage unit that stores time series of feature vectors and silent interval points of input patterns and all standard patterns. Reference numeral 6 denotes a back pointer storage unit which stores the intersection points of silent interval points passed before reaching each point when calculating the cumulative distance at each point. 7
The recognition determination unit uses the standard pattern that provides the minimum distance among the distance values obtained by the similarity calculation unit 4 as the 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, audio is input from the microphone 1, and the audio analysis unit 2
The input audio signal is analog-to-digital converted, and a time series of audio feature vectors (for example, 10th order linear prediction coefficients) and an energy series are obtained and stored in the storage unit 5. Next, the silent section detecting section 3 detects, as a silent section, a section in which the energy value is below a predetermined threshold and exceeds a certain time TO from the energy sequence stored in the memory by the speech analyzing section 2. The number of sections and the frame position of the rise of energy following each silent section (hereinafter, this frame position will be referred to as Q point) are stored. Note that the time series of the feature vectors of each standard pattern, the number of silent sections of each standard pattern, and the frame position of the rise of the subsequent energy are stored in the storage unit 5 in advance.
It is assumed that it is stored in . Next, in the similarity calculation unit 4,
The input audio pattern is subjected to pattern matching (for example, DP matching method) in which the matching path is constrained by Q points detected by the silent section detection section 3 between each standard pattern. For simplicity, the pattern matching operation will be described below for the case where there are two Q points in each of the input pattern and the standard pattern.

Here, the start point and end point of the input pattern are respectively Q+a
(= 1 ), Q10, and the start and end points of the standard pattern are QR9 (= 1 ) and QR3, respectively. Also,
From the Q point of the input pattern and standard pattern, from the one closest to the starting point, Q ++s Q 12 and Q R+, QR2, respectively.
Let the intersection of each Q point on the two-dimensional plane be (QI
IIQRI) (QRIQ-2) (Q-2Q-1) (QR
20R2).

These conditions are shown in FIG. Also, for simplicity, we will not consider restrictions due to the matching window.

Pattern matching is performed by restricting the matching path to the intersections of the Q points detected on the input pattern and the standard pattern, but considering the case where the Q points are detected incorrectly, as shown in Figure 2. It is necessary to consider three types of routes: ■■■. Here, we consider the following equation as a recurrence equation for pattern matching.

g(1,1):d(1,1) --
---(1) g(1+j)=mln(g(11,J)
+g(1-1+j-ILg(1-1,j-2))+d(
1, j) -----(2) g(1, j)
=g(1+1)=co(1,J=2.3.* e e
e) Here, g(i, j) and d(i, j) are each (
t, j) and the distance between vectors.

Also, the back pointer at point (t, j) is b(i, j)
and the initial value is b(1,1)E(1,1).

First, set i''Lj=1 and calculate the cumulative distance g according to the above recurrence formula while incrementing j, and then calculate j''QIIE
When processing is completed, i is incremented and processing continues in the same manner. The back pointer at each point (i, j) is determined as follows.

! =Q+ and j=QR+ (However, I=1
．． 2) b (i+ j ) E (Q+b+ QR
I) When i f−Q Ikl or j≠QRI (however, I:l, 2) (i, j) is Q, ,<i≦
When Q+ is 4+ and QRI<j≦QR1+1, the content of the back pointer corresponding to gO, which is the minimum value of the first term on the right side of recurrence formula (2), is expressed as b(i
, j). However, here b(i.

The points (Qll, QR,) that can be taken as j) are (m=
k and n=1) or (n=1 and m=k).

The back pointer b(i, j) thus obtained is stored in the back pointer storage section 6. The cumulative distance at the point (Q10.QR3) finally obtained in this way is the degree of similarity when the most optimal limited route is selected considering all three types of routes in FIG. 2.

After all similarities with each standard pattern are calculated,
The recognition determining unit 7 determines the standard pattern with the highest degree of similarity as the recognition candidate speech and outputs it to the outside.

As described above, according to this embodiment, the silent section detection unit 3 detects silent sections of audio, and the similarity calculation unit 3 calculates the similarity between an input pattern and a standard pattern by pattern matching in which a path is restricted by the intersection of the silent sections. By providing a degree calculation unit 4 and a back pointer storage unit 6 that stores the intersection points of Q points passed in the past, it is possible to consider all possibilities of route limitation in one pattern matching, and calculate It is possible to prevent erroneous recognition without increasing time, and to reject as much as possible non-target audio input.

Effects of the Invention As described above, the present invention includes a silent section detecting section that detects a silent section in a voice pattern and stores its position, and a pattern that is restricted to pass through the intersection of the silent section points of the input pattern and the standard pattern. By providing a similarity calculation unit that uses a back pointer to calculate the cumulative distance and memorizes the intersection points of optimal silent section points passed in the past, pattern matching corresponding to the energy envelope is possible. In addition, since it considers how to match all silence interval points in one pattern matching calculation, it reduces misrecognition between similar words without increasing processing time, and eliminates non-target words. It is possible to provide a speech recognition device that can reject as much as possible even when input.

[Brief explanation of the drawing]

Fig. 1 is a block configuration 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. Detection unit, 4e similarity calculation unit, 5... storage unit, 6... Φ back pointer storage unit, 7... @ recognition determination unit.

Claims (2)

[Claims] (1) a speech analysis unit that outputs a time series of feature vectors including energy sequences from input speech; a silent section detection unit that detects silent portions in the speech pattern from the energy series output from the speech analysis unit; A pattern matching distance calculation that limits a pattern matching path at the intersection of the silent interval points of the input pattern obtained by the silent interval detection unit and each standard pattern stored in the dictionary in advance is performed by taking into account the error in detecting the silent interval points. A speech recognition device comprising: a similarity calculation unit that calculates a similarity between the two by calculating a distance on a matching path limited by the intersections of all possible silent interval points. (2) A claim characterized in that the similarity calculation unit has a back pointer that stores which intersection of silent interval points the matching path has passed before reaching that point when calculating the cumulative distance at each point. Speech recognition device according to item 1