JPH0458635B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Application Field] The present invention relates to a method for extracting a standard phoneme pattern in a speech recognition device. [Prior art] Speech recognition is a process of acoustically analyzing speech to extract the linguistic features contained therein, and converting this into a display of linguistic symbols corresponding to the speech.In principle, there are two types of speech recognition: method is known. That is,
One method is to memorize in advance a standard pattern related to the linguistic features contained in speech, compare this standard pattern with the speech input to check for similarities, and based on that similarity, the input speech input is set as the standard pattern. This is a method of recognizing and determining whether or not it matches a pattern. Another method is to repeatedly judge phoneme symbols based on the acoustic analysis results of the speech input without using the standard pattern described above, and finally make a recognition judgment as a language. It is. Among the above two methods, the former method using a standard pattern generally gives good recognition results, and for example, the method shown in FIG. 8 is used to recognize words from voice input. In FIG. 8, after acoustic analysis is performed using the frequency spectrum envelope of the input voice input and two acoustic features of the driving sound source through correlation analysis, etc., phoneme recognition is performed using a phoneme standard pattern created in advance. In this phoneme recognition, the input acoustic features are expressed as a series of phoneme symbols, and this series of phoneme symbols is recognized as a word using a word dictionary created in advance, and the recognized word is the linguistic symbol of the word. is output in the form of [Problems to be Solved by the Invention] As mentioned above, when using phonemes as the basic unit of recognition in continuous speech recognition, it is necessary to cut out a standard phoneme pattern from the registered word speech in advance, and this phoneme cutting is done as follows: Conventionally, inspections were conducted by experts in speech information processing, which took a long time and was extremely inconvenient. [Object of the Invention] An object of the present invention is to solve the above-mentioned problems and to provide a method for extracting a standard phoneme pattern that can mechanically and quickly extract a standard phoneme pattern from a word sound without manual intervention. It is in. [Structure of the Invention] The present invention stores in advance a word network having a plurality of transition paths with phoneme boundary symbols as nodes for each word uttered by a plurality of speakers, while A phoneme boundary symbol string and a speech analysis parameter sequence are extracted, and when the phoneme boundary symbol string of the input word speech matches at least one transition path of the word network stored in the storage means, the It is characterized by extracting a parameter sequence as a standard phoneme pattern for speech recognition. [Example] Fig. 1 is a block diagram of a phoneme standard pattern cutting device which is an example of the present invention. It is characterized by the use of a word network whose nodes are phoneme boundary symbols. In FIG. 1, first, the word speech for registration X(t) is input to the speech analysis section 1, and from the speech input X(t),
Autocorrelation coefficient R(t) and its change R′(t), power P
(t) and its change P′(t), and cepstral coefficient c
(t) is calculated. Here, the frame period of the audio input is assumed to be 8 msec, and the above t is the t of the audio input.
Represents the number frame. FIG. 2 is a block diagram of the speech analysis unit 1 shown in FIG. 1. In FIG. The sampled value S(t) is output to the autocorrelation coefficient calculating section 12 and the power calculating section 13. The sampling circuit 11 of this embodiment performs sampling 256 times per frame, and hereinafter, each sampling value is expressed as S(t)i, 1≦i≦256 (1). In the autocorrelation coefficient calculation unit 12, from the input sampling value S(t), the third
Based on the processing flow in the figure, the autocorrelation coefficient R of the following formula
After (t)i is calculated, it is output to the linear prediction coefficient calculation section 14 and the phoneme classification section 2. R(t)i=1/256 _256-i 〓 ^k=1 S(t)k ・S(t)k+i, 1≦i≦24 ……(2) Here, the subscript i is the autocorrelation coefficient R(t ) represents the order of the linear predictive analysis integer A described below.
Each subscript i of (t)i and cepstral coefficient c(t)i also represents the order. In the flowchart of FIG. 3, S(I) represents the sampled value S(t)i, and R(I) represents the autocorrelation coefficient R(t)i. In the linear prediction coefficient calculation unit 14, a linear prediction analysis coefficient A(t)i is calculated from the input autocorrelation coefficient R(t)i by a known linear prediction analysis method based on the processing flow shown in FIG. After that, it is output to the cepstral coefficient calculation section 15. In the cepstrum coefficient calculation unit 15, a cepstrum coefficient c(t)i is calculated from the input linear prediction analysis coefficient A(t)i using the following equation, and is output to the phoneme extraction unit 4 and the cepstrum change calculation unit 16. c(t)i=-A(t)i-1/i _i-1 〓 ^k=1 k・c(t)k・A(t)i−k, 1≦i≦24 ……(3) However , (3), the first-order cepstrum coefficient c(t) ₁ is expressed by the following equation. c(t) ₁ = -A(t) ₁ ...(4) Furthermore, the cepstrum change calculation unit 16 calculates the change in cepstrum coefficient c′(t )i is calculated and output to the phoneme boundary detection section 3. c'(t)i=|c(t-4)i-c(t)i| ...(5) On the other hand, in the power calculation section 13, from the input sampled value S(t)i, the following equation is calculated. After the power P(t) is calculated based on the power P(t), it is output to the phoneme classification section 2 and the power change calculation section 17. P(t)=1/256 ₂₅₆ 〓 ⁱ⁼¹ |S(t)i| ² ...(6) Next, in the power change calculation section 17, the input power P(t) is calculated based on the following formula. power change
P'(t) is calculated and output to the speech boundary detection section 3. P'(t)= ₇ 〓 ^j=1 (j-4)・P(t-7+j) ...(7) Figure 5 is an area table for phoneme classification in the phoneme classification section 2 in Figure 1. Yes, the horizontal axis X is −log(1
-R(t) ₁ ), and the vertical axis Y is logP(t). Here, R(t) ₁ is 1 of the tth frame as described above.
The following autocorrelation coefficient is: In FIG. 5, an area where Y is less than a predetermined boundary value _Y1 is a silent part (.). Furthermore, in an area where Y is greater than or equal to the predetermined boundary value _Y1 and less than or equal to the predetermined boundary value _Y2 , and where X is less than the predetermined boundary value _X1 , a silent part (F) is formed, _and A region where X is above and below a predetermined boundary value X ₂ is a vowel part (V), and a region where X exceeds a predetermined boundary value X ₂ is a nasal part (N). Furthermore, the region where Y exceeds Y ₂ and where Y<-m ₁ (X-X ₁ )+Y ₂ ...(8) is a silent part (F), and Y≧-m ₁ (X- The region where X ₁ )+Y ₂ ...(9) and Y≧m ₂ (X-X ₂ )+Y ₂ ...(10) is the vowel part (V), and Y<m ₂ (X-X ₂ ) + Y ₂ ...(11) The region is the nasal part (N). Here, m ₁ and m ₂
is a positive predetermined value. In the phoneme classification section 2, the input power P
(t) and the autocorrelation coefficient R(t), the general characteristics of each frame of speech input are determined by the phoneme classification symbol based on Figure 5.
It is output to the phoneme boundary detection unit 3 in the form of ph(t). In addition,
Table 1 shows the output phoneme classification symbol ph(t) and the properties it should represent. Next, the phoneme boundary detection unit 3 uses the input power change P'(t), cepstral coefficient change C'(t)i, and phoneme classification symbol ph(t) to calculate the , the phoneme boundary number bd(t) in Table 2 is detected,
It is output to the phoneme extraction section 4. Note that in Table 2, T ₁ , T ₂ and T ₃ are predetermined threshold values. In this phoneme boundary detection unit 3, if the interval between boundary numbers is within a predetermined threshold T ₄ frames, a phoneme boundary number with a high priority as shown in the following formula
bd(t) is output. High priority >>>>>Low priority...
(12) Figure 6 shows the phonological classification symbol sequence ph(t) and the boundary number sequence bd(t) when three speakers uttered “Asahi”.
It is a figure showing an example. As described above, one word section can be described by a boundary symbol string bd(t) that starts with a boundary symbol and ends with a boundary symbol. When the boundary symbol string bd(t) in FIG. 6 is expressed as a word network with boundary symbols as nodes, the result is as shown in FIG. 7. However, on the branches between nodes, the phonemes that exist in that section are indicated with serial numbers above the nodes. As shown in Fig. 7, in a word network for one word created by multiple speakers, multiple transitions occur because the boundary symbol string bd(t) differs depending on the speaker. A road exists. In Figure 1, numeral 5 is a word network table (ROM), which analyzes the audio data of words for phoneme segmentation uttered by many speakers in advance, and creates a word network for each word as shown in Figure 7. make,
Write it in word network table (ROM) 5. In order to store this network in the memory (ROM), a list representation as shown in Table 3 is used, and one branch is represented by 6 words of node information as shown in Table 3. Table 3 shows the meaning of each word in the node information, and Table 4 shows the extraction positions of phonemes in each branch and their symbols. In Table 3, the branching condition (shortest) is the minimum frame interval until a boundary symbol that satisfies the branching condition arrives, and the branching condition (longest) is the frame interval until a boundary symbol that satisfies the branching condition arrives. This is the maximum value of the interval. In the example in Table 3, if the boundary symbol comes within 5 frames or more and within 15 frames, the node number 4
The cepstral coefficient c in the center frame of the section connecting the current node and the branch destination node
This means cutting out (t) as a standard pattern of the phoneme /a/. The phoneme extraction unit 4 reads out the word network corresponding to each word for phoneme extraction from the word network table (ROM) 5, and analyzes the voice input for registration to generate a boundary symbol string output from the phoneme boundary detection unit 3. bd(t) is input. First, starting from the boundary symbol that is the first node, if the branching condition in the node information is satisfied, the pointer provided in the phoneme extraction section is moved to the next node,
Repeat this action. Based on the input boundary symbol string bd(t), the pointer can transition according to the word network stored in the word network table (ROM) 5, and can transition to the boundary symbol representing the end of the word. Only when it is possible to segment the phoneme, it is assumed that the phoneme has been successfully segmented, and the cepstral coefficient c(t ₀ ) in the frame corresponding to the extraction position t ₀ of the node information written in the word network table (ROM) 5 is extracted for each phoneme. , and its coefficient c(t ₀ ) is stored in the phoneme standard pattern table (RAM) 6 as a standard phoneme pattern. As explained above, by analyzing the audio data of words for phoneme segmentation uttered by many speakers in advance,
A word network with phoneme boundary symbols as nodes, as shown in Figure 7, is written in the word network table (ROM) 5 in the form of the node information in Table 3, where the branches between each node are represented by 6 words. , the boundary symbol string bd(t) analyzed from the registration voice input X(t) is compared with the word network written in the word network table (ROM) 5, and if there is a matching transition path, The cepstral coefficient c(t ₀ ) in the frame corresponding to the extraction position t ₀ of the node information written in the word network table (ROM) 5 after determining that the phoneme has been successfully segmented is extracted as a phoneme standard pattern for each phoneme. be able to.

【table】

【table】

【table】

【table】

〔Effect of the invention〕

As detailed above, a word network uttered by multiple speakers is analyzed in advance, and a word network having a plurality of transition paths with phoneme boundary symbols as nodes is stored in a storage means for each word, and the word network is inputted. the input phoneme boundary symbol string matches at least one transition path of the word network stored in the storage means; In this case, since the parameter series can be extracted as a standard pattern of phonemes for speech recognition, the standard pattern of phonemes can be extracted mechanically and quickly from word speech without human intervention.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a phoneme standard pattern extraction device which is an embodiment of the present invention, and FIG.
3 is a flowchart showing the processing of the autocorrelation coefficient calculation section of FIG. 2, and FIG. 4 is a flowchart of the processing of the linear prediction analysis coefficient calculation section of FIG. 2. , Figure 5 is a diagram showing the classification area in the phoneme classification section of Figure 1, and Figure 6 is a diagram showing the phoneme classification symbol string and boundary number string when three speakers utter "Asahi". , Figure 7 is the 6th
FIG. 8 is a block diagram showing a conventional speech recognition method.

Claims (1)

[Scope of Claims] 1. A word network having a plurality of transition paths with phoneme boundary symbols as nodes for each word uttered by a plurality of speakers is stored in advance in a storage means, and the phoneme boundaries of input word sounds are stored in advance in a storage means. A symbol string and a speech analysis parameter series are extracted, and when the phoneme boundary symbol string of the input word voice matches at least one transition path of the word network stored in the storage means, the parameter series is extracted. A method for extracting a standard phoneme pattern, characterized in that the standard phoneme pattern is extracted as a standard phoneme pattern for speech recognition.