JPS6147999A - Voice recognition system - 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]

<Industrial Application Field> The present invention relates to an improvement of the voice B-weaving method in a voice input device used in a voice input type document processing device or the like. <Prior art> Human speech is decomposed into smaller syllable units (hereinafter referred to as syllable segmentation), and a time series of syllable candidates consisting of a plurality of syllable candidates identified for each syllable ( Conventionally, when the candidate syllable string obtained by combining candidate syllables with high identification accuracy based on the syllable lattice (hereinafter referred to as the syllable lattice) is compared with a dictionary and the obtained candidate is used as the recognition result for the input speech, the syllable lattice used in the above syllable segmentation It is common practice to uniquely determine boundaries. However, with this type of processing, as the input speech becomes more continuous and approaches conversational speech, syllable boundaries become unclear and difficult to detect, making recognition difficult. The inventors focused on this point, and in order to improve the accuracy of syllable boundary detection, the syllable segmentation algorithm does not uniquely determine syllable boundaries, but instead estimates the speech rate. proposed to improve the recognition rate by utilizing this estimated value, but this still causes errors in syllable segmentation. <Problem to be solved by the invention, α> The present invention solves the following problems. , syllable segmentation is still difficult even when using speech rate.
This was done with the aim of solving the problem of error 1) and providing a speech recognition method with a good recognition rate. Means for Solving Problems> In order to achieve the above-mentioned object, the present invention determines the identification accuracy of a candidate syllable string using both the identification distance of each syllable and the reliability based on the utterance rate. It is characterized by the fact that In the method of the present invention, syllable boundaries are not determined uniquely, but two factors are used: syllable identification distance and reliability based on utterance speed, and syllable boundaries are detected as multiple candidates and We perform syllable segmentation that allows the existence of multiple syllable intervals that compete with each other. Then, using a competitive syllable candidate lattice obtained by identifying syllables for each syllable interval candidate, a candidate syllable string is created with priority given to syllable intervals with high reliability of syllable interval detection. <Example> The present invention will be specifically described below with reference to an example shown in the drawings. FIG. 1 is a diagram showing the structure of a voice input device embodying the present invention. The input speech is first subjected to acoustic analysis in the speech analysis section 1, and sent to the silent section detection section to detect silent sections.Furthermore, the phrase boundary detection section 5 detects whether or not it is a silent section as a phrase boundary. It will be done. On the other hand, the voiced section detection section 3 detects a voiced section, that is, a cluster of sounds, and the syllable boundary candidate detection section 6 detects power changes, spectrum changes, phoneme changes, etc., and the utterances stored in the speech rate storage section 4. A result of 0 or more in which syllable boundary candidates are given using velocity information is a speech
- It is integrated in the syllable interval candidate detection unit 7 together with the clear syllable boundary information obtained by the analysis unit 1 to obtain valid syllable interval candidates in order of reliability. These syllable interval candidates are identified by the syllable identification unit 8 through pattern matching with syllable standard patterns that are stored in advance in the pattern memory 12, and as will be described later, candidate syllables are sorted in descending order of identification distance. Output. The time series of syllable candidates having the identification distance thus output is the syllable candidate lattice (b). On the other hand, the syllable interval candidate confidence calculation unit □9 uses the speech rate information stored in the speech rate storage unit 4 to calculate the reliability of the syllable interval candidate according to equation (1) described later. A syllable interval sequence is estimated. The candidate syllable sequence creation unit 10 determines the weighting coefficient K of each syllable interval for other syllable interval series based on this optimal syllable interval series, and calculates the identification distance and the reliability of interval detection according to equation (2) described later. candidate syllable strings are created sequentially in descending order of reliability S. The candidate syllable string is sequentially subjected to language processing including dictionary checking using the dictionary 13 in the language processing unit 11, and an interpretable candidate string is output as a recognition result (c). And the speaker (
The length of each syllable in the syllable block 'fJJ series constituting the candidate string determined by the operator) is stored in the speech rate storage unit 4 as speech rate information together with the previous average syllable length value, and is used for subsequent processing. I get frowned upon. Next, the process of determining syllable intervals will be described.

[1] Calculation of reliability of syllable interval The reliability of a syllable interval candidate based on the speaking rate is determined by the deviation from the average value of the syllable length. Therefore, when using a voice input device, the user first inputs a standard sentence whose number of syllables is known, and the average syllable length is estimated. The average syllable length is 1 of the first voiced interval of a sentence with n syllables.
If one duration time is L(i), it is calculated as follows. Therefore, the reliability DIID21...*Dx*... for multiple syllable interval sequences within the same time is as follows (1)
It can be found using a formula. However, X(i): the first syllable length X of the syllable interval series X
: Number of syllable intervals L in syllable interval series (i)
−L I /L. - DX: Reliability of syllable interval series X The syllable interval with the minimum reliability determined by this equation (1) is the optimal syllable interval series.

[2] Creation of a candidate syllable string A syllable candidate lattice in which multiple syllable interval candidates are detected within the same time and is in a competitive state is evaluated using both the syllable identification distance and the reliability of the syllable interval described above. S is determined, and candidate syllable strings are created in descending order of this value. The syllable identification distance is determined as follows based on the above-mentioned optimal syllable interval series. Now, a syllable interval candidate as shown in Fig. 2 has been detected, and B-
Suppose that the C series and the D series are competing with each other. Here, La1aj: Discrimination distance between the syllable length of syllable interval candidate A and the i-th syllable candidate Lb, bj: Discrimination distance between the syllable length of syllable interval candidate B and the j-th syllable candidate LC+Ck: The syllable length of syllable interval candidate C and the i-th syllable candidate. Assuming that the identification distance Ld, dl of the syllable candidate is the syllable length of the syllable interval candidate and the identification distance L of the first syllable candidate is the average syllable length, the reliability of the competing sequences is as follows: Reliability of the B-C sequence Dbc= id(Lb, L)+d(Lc, L)
l/ 2D sequence reliability Dd=dCLd,L), and the evaluation value of the competitive part is 5bc=(bi+ck)X Kbc+ DbcScl=
It is determined by dlXKd+Dd. However, Kbj and Kd are weighting coefficients for the identification distance based on the optimal syllable interval sequence, and when B-C is the optimal sequence (that is, when Dbc<Dd), Kb
c”1.Kd=2, and when D is the optimal sequence (Dd<
Dbc) is Kd=1-Kbc=1/2. From the above results, the overall evaluation value is: The smaller one of S abc and S ad is set as the evaluation value S, and candidate syllable strings are created in the order of combinations with smaller values. The above procedure will be specifically explained using the input voice "House" as an example. The third factor is a diagram showing the state of syllable boundaries when "house" is input. That is, there are four syllable section candidates A, B, C, and D, and each candidate B is a competing portion. The syllable lengths of C and D are each Lb=12. Lc=27. L
Assume that d=39 and the average syllable length U=23. The reliability of each syllable interval is as follows: For B: d(Lb, U) = 0.48 For C: d(LcJ:) = 0.17 For D: d(
Ld, ]:)=0.70. Therefore, the reliability of each syllable interval sequence is the B-C sequence: Dbc = (d (Lb, U) + d (Lc, t,), /
2 =0.32D series: Dd=d(Ld,'L:) =0.70
Therefore, it is estimated that the B-CilS sequence is the optimal syllable interval sequence. If the reliability is normalized for this B-C sequence, D'bc''O D'd = Q, 38. Figure 4 is a syllable candidate lattice showing the identification distance of each candidate blue line. Using each identification distance and the above-mentioned reliability (after normalization), the evaluation value of each candidate syllable string is calculated according to formula (2) as 5abc=ai+(bj+ck+D'bc)Sacl
=ai+(dlX 2 +D'd), and candidate columns are created one after another in the order of combinations that give the smallest value. The candidate syllable strings created in this way are subjected to language processing, and the recognition results are output as shown in the attached table. According to this result, the first candidate is "Ieo" from the ABC series.
It can be seen that the correct results are obtained. <Effects of the Invention> As is clear from the above description, according to the present invention, 1. Since syllable segmentation is performed using both the identification distance and utterance speed of each syllable, errors in segmentation are reduced, and subsequent This makes it possible to increase the probability of correct correction through language processing.

[Brief explanation of drawings]

Figure 1 is a block diagram showing the configuration of a voice input device implementing the present invention, Figure 2 is a diagram showing a syllable candidate lattice, same as above, and Figure 3 is an example of syllable boundaries of input speech, same as above. FIG. 4 is a diagram showing a specific example of a syllable candidate lattice. 1... Speech analysis section 4... Speech rate storage section 7... Syllable section candidate detection section 8... Syllable identification section 9... Syllable section candidate reliability calculation section 10... Candidate syllable Column creation unit 11...language processing unit

Claims (1)

[Claims] The input speech is broken down into syllables, each syllable is identified, and candidate syllable strings corresponding to words and phrases are sequentially created based on the time series of multiple syllable candidates in descending order of identification accuracy to generate word sounds and phrases. A speech recognition method, in a speech input device for recognizing speech, characterized in that the identification accuracy of a candidate syllable string is determined using both the identification distance of each syllable and the reliability based on the utterance speed.