JPH04204899A - Speech recognition device - Google Patents

Abstract

PURPOSE:To obtain good degrees of section similarity and to provide a word recognition system as well as to improve a recognition rate by subjecting the degree of frame similarity to weighting computation in a time direction, then determining the degree of section similarity. CONSTITUTION:A means 2 for calculating the degree of the frame similarity outputs the degree of the frame similarity obtd. by calculating the degree of similarity between the characteristic parameter time series obtd. by a speech analyzing means 1 and the standard parameter prepd. for each of the phonemes stored in a phoneme standard pattern 3. A means 4 for computing weighting obtains the weighted degree of the frame similarity by referencing the weighting function stored in a means 5 for storing the weighting functions. A means 6 for calculating the degree of the section similarity calculates the quantity of the degree of the section similarity with all the phonemes and a means 7 for calculating the degree of the word similarity obtains the degree of the word similarity by most adequately adding the degree of the section similarity in accordance with the array of the phonemes concerning the respective words obtd. by referencing a word dictionary 8. The recognition rate is improved in this way.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device that automatically recognizes human speech.

2. Description of the Related Art Conventionally, as a method for realizing a device that recognizes human speech, there is a method in which phonemes and syllables are used as basic units of recognition. This method has the advantage that it is easy to change the word dictionary because it does not require voice registration. Hereinafter, as a conventional technique, a method using phonemes as the basic unit of recognition, a method of obtaining the similarity of phonemes for each frame and then integrating them to obtain a final recognition result will be described.

FIG. 2 shows the configuration of a conventional speech recognition system.

In FIG. 2, input speech is input to speech analysis means 11. In FIG. This part, that is, the speech analysis means 11 analyzes the input speech and outputs a time series of feature parameters representing the characteristics of the speech to the frame similarity calculation means 12 for each frame, which is the basic unit of analysis. The frame similarity calculation means 12 calculates the similarity between the feature parameter time series obtained by the preceding speech analysis means 11 and the standard pattern prepared for each phoneme stored in the phoneme standard pattern 13. The frame similarity obtained by doing this is output to the section similarity calculation means 14.

The section similarity calculating means 14 at the next stage calculates a quantity called "section similarity" which is the similarity of phonemes for partial sections in the input speech for all phonemes.

The section similarity is a value obtained by accumulating the frame similarity for a partial section and normalizing it by the section length. The word similarity calculation means 15 obtains a word similarity by optimally adding up the section similarities according to the phoneme arrangement for each word obtained by referring to the word dictionary 16.

Next, a method of calculating word similarity using interval similarity will be explained.

The phoneme sequence of the word in the word dictionary is p(1)I)(2)l
)(3)...p (N). The lower limit of the duration of the i-th phoneme p(i) in a word is l mi
n (i), and the upper limit value is 1 max (+). The frame similarity in the j-th frame of phoneme P(1) is R
(l, j), the cumulative similarity of phoneme p(1) at frame number j is 5CORE (i, j). Further, the interval similarity of phoneme p in the partial interval of final frame number J and length L is assumed to be SIM(p, j, L). Also 5 CORE
When (i, j) is obtained, the length (number of frames) from the start of the input audio to J is 1en(i, j). Furthermore, the symbols are defined as follows.

SS (i): First frame number of the starting range of p (i) SE (i): Last frame number of the starting range of p(i) ES (i): First frame of the ending range of p (i) Number EE (i): Last frame number of the terminal range of p (i) The procedure for calculating word similarity is shown below.

(1) Initialization SS (1) ← First frame number E of the range where the start of the word exists (1) ← Last frame number of the range where the start of the word exists 5CORE (0,・)-〇len (0,・) =0 (2) Execute (3) and (4) from i=1 to iN.

(3) Determination of phoneme ending range ES (j) ←SS (j) +1m1n(i)EE
(i)←SE (i)+1max(1)(4) Calculation of cumulative similarity in phoneme end range j=ES (Execute the following recurrence formula from (i) to EE(i).

5CORE (i, j) − MAX[+αxSIM(p (i), j, L) 1 mi
n(i)≦L≦lmax(i)
\However, α, β, and γ are weighting coefficients for time normalization.

α=L β=1en(i-1,j-L) γ=α+β (5) Determination of word similarity Let the value R obtained by the following equation be the word similarity.

R=MAX I 5CORE (N, j) 1''C Existence area of word end (6) End Find word similarity R for all words stored in the word dictionary according to the procedure described above.

Finally, the word determining means 17 in FIG. 2 obtains the word giving the maximum word similarity value as a recognition result.

Problems to be Solved by the Invention As described above, in the conventional speech recognition method, word similarity is calculated based on the cumulative value of phoneme similarity for each frame. However, this method, ie, the method of multiplying and using this as the similarity of phonemes in the section, has a problem for the reasons described below.

The degree of phoneme similarity for the actual phoneme existence interval does not necessarily match the value obtained by simply accumulating the phoneme similarity for each frame for the aforementioned existence interval. In other words, the standard pattern of phonemes provided for calculating frame similarity is:
It is common to create a time series of feature parameters over a wide interval in the time axis direction, including the vicinity of a frame (referred to as an epoch frame) in which phoneme features become noticeable, from audio data.

If the standard pattern with a time width created using this method is shifted one frame at a time and the phoneme similarity is calculated for each frame, the frame similarity in the epoch frame of the phoneme is clearly meaningful, but There is no guarantee that the frame similarity value in is a meaningful value.

However, in the conventional method, frame similarity in epoch frames and frame similarity in non-epoch frames are treated equally to calculate interval similarity, which causes a problem in the quality of interval similarity, which is one of the reasons for the decline in recognition rate. It became.

In view of the above problems, it is an object of the present invention to improve the recognition rate by obtaining a high-quality section similarity and realizing a word recognition system.

Means for Solving the Problems To achieve this object, the present invention provides a speech analysis means for analyzing input speech frame by frame, which is a unit of analysis, and obtaining feature parameters, and a method for expressing the features of the basic unit of recognition. A standard pattern storage means for storing a standard pattern, a frame similarity calculating means for calculating the degree of similarity between a feature parameter and a standard pattern for each frame, and a pronunciation of a word described by a symbol representing a basic unit of recognition. weighting function storage means for storing or generating a weighting function in the time direction; and weighting calculation using the weighting function for the frame similarity calculated by the frame similarity calculation means. The word similarity is calculated by accumulating the frame similarity for each basic unit of recognition obtained by the weighting calculation according to the notation in the word dictionary.

According to the above configuration, the present invention calculates the interval similarity by accumulating the values after weighting the frame similarity in the time direction, thereby obtaining a higher quality interval similarity than the conventional method. By using this interval similarity to implement a word recognition system, it is possible to improve the recognition rate.

EXAMPLE Hereinafter, an example 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, input speech is input to speech analysis means 1. In FIG. This part, that is, the voice analysis means 1 analyzes the input voice, and calculates the time series of feature parameters representing the characteristics of the voice for each frame, which is the basic unit of analysis, into the frame similarity calculation means 2.
Output to.

The frame similarity calculation means 2 calculates the similarity between the feature parameter time series obtained by the preceding speech analysis means 1 and the standard pattern prepared for each phoneme stored in the phoneme standard pattern 3. The frame similarity obtained by this method is output.

Note that the phoneme standard pattern 3 stores standard patterns representing characteristics of basic units of recognition such as phonemes and syllables.

The weighting calculation means 4 performs a weighting calculation on the frame similarity obtained in the previous stage by referring to the weighting function stored in the weighting function storage means 5 to obtain a "weighted frame similarity".

The weighting calculation is performed by setting a partial section of the input speech section. Therefore, a weighting function is prepared for each subinterval length, and weighs the degree of contribution of frame similarity at the center of the subinterval and its vicinity, and weighs the degree of contribution of frame similarity at both ends of the subinterval and its vicinity. Make it smaller.

The section similarity calculation means 6 at the next stage calculates a quantity called "section class lowness" which is the similarity of phonemes for partial sections in the input speech for all phonemes.

The section similarity is a value obtained by accumulating the above-mentioned "weighted frame similarity" for a partial section and normalizing it by the section length.

The word similarity calculation means 7 obtains the word similarity by optimally adding up the section similarities according to the phoneme arrangement for each word obtained by referring to the word dictionary 8.

Next, a method of calculating word similarity using interval similarity will be explained. The phoneme arrangement of the word in the word dictionary 8 is p
(i) p (2) p (3)...p(N)
Suppose that The lower limit of the duration of the first phoneme p(1) in a word is 1 min (i), and the upper limit is l ma
Let x(1) be. Let the frame similarity of phoneme P(1) in the j-th frame be R(i, j), and the cumulative similarity of phoneme p(i) in frame number j be 5CORE(i, j). Also, let SIM(p, j, L) be the interval similarity of phoneme p in a partial interval with final frame number J and length. Also 5CORE (i.

'') is obtained, the length (number of frames) from the start of the input audio to j is defined as Ien(i, j). Furthermore, the symbols are defined as follows.

SS (i): I) First frame number of the starting range of (i) 5E (i): Last frame number of the starting range of p(1) ES (i): First frame number of the ending range of p (i) The procedure for detecting the last frame number word similarity in the end range of EE (i):p(j) is shown below.

(1) Initialization SS (1) ← First frame number E of the range where the start of the word exists (1) ← Last frame number of the range where the start of the word exists 5 CORE (0, ・)=O 1en (0+ ・)= 0 (2) Execute (3) and (4> from i=1 to i=N.

(3) Determination of phoneme ending range ES (i) ←SS (i)+1m1n(i)EE
(i)←SE (i)+1max(1)(4) Calculation of cumulative similarity in phoneme end range j=ES (Execute the following recurrence formula from (i) to EE(1).

5CORE (i, j) - MAX mouth (αXSIM (p(i), j, L)l mi
n (i)≦L≦1 max (i)+βxscO
RE (i-1,j-L)l/γ] where α, β, and γ are weighting coefficients for time normalization.

α=L β=len(i-1,j-L) γ=α+β (5) Determination of word similarity Let the value R obtained by the following equation be the word similarity.

R=MAXISCORE (N, j) 1jC Existence region of word end (6) End Find the word similarity R for all words stored in the word dictionary according to the procedure described above.

Finally, the word determining means 9 in FIG. 1 obtains the word giving the maximum word similarity value as a recognition result.

In the above embodiments, a specific example has been described in which the basic unit of recognition is a phoneme, but the present invention is of course applicable to cases where the basic unit of recognition is other than a phoneme, such as a syllable.

Effects of the Invention As described above, according to the present invention, by calculating the interval similarity after performing a temporal weighting operation on the frame similarity, the value of the frame similarity in meaningful frames is emphasized. It is possible to obtain a high-quality section similarity, and by implementing a word recognition system using this section similarity, it is possible to improve the recognition rate.

[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 a block diagram of a conventional speech recognition device. ]...Speech analysis means, 2...Frame similarity calculation means, 3...Phoneme standard pattern, 4...Weighting calculation means, 5...Weighting function storage means, 6...Section similarity Calculation means, 7... Word similarity calculation means, 8...
Word dictionary, 9... word determination means.

Claims (1)

[Scope of Claims] Speech analysis means for analyzing input speech frame by frame, which is a unit of analysis, to obtain feature parameters; and standard pattern storage means for storing a standard pattern representing the features of the basic unit of recognition; a frame similarity calculation means for calculating the similarity between the parameters and the standard pattern for each frame; a word dictionary storing information regarding the pronunciation of words described by symbols representing the basic units of recognition; weighting function storage means for storing or generating a direction weighting function; and weighting calculation means for performing a weighting operation using the weighting function on the frame similarity calculated by the frame similarity calculation means, A speech recognition device characterized in that a word similarity is calculated by accumulating the frame similarity for each basic unit of recognition obtained by calculation according to the notation in the word dictionary.