JPS60237496A - Voice recognition equipment - 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] 辣]L年ノ□ The present invention relates to a human speech recognition device using vector demonization.

1. Prior Art Figure 1 is a basic circuit diagram of a speech recognition device. In the figure, 1 is a microphone, 2 is an analysis section, 3 is a changeover switch, 4 is a standard pattern section, 5 is an input speech pattern section, and 6 is a basic circuit diagram of a speech recognition device. Distance calculation section, 7 is minimum value detection section, 8 is recognition! flit i:'i
In the disaster area, the pattern matching part is formed by the 2-point scale part 6 and the minimum value detection sound 57. In Fig. 1, first, the sound coming from the microphone 1 is analyzed and its sound pattern is In a system for a specific speaker, before recognition, the analysis results of each word to be recognized by that speaker are registered as standard patterns. Compare the parameters of the standard pattern of the target word and the input speech pattern,
Select the closest word to be recognized, that is, the word with the smallest distance. Note that in the case of unspecified speakers, a standard pattern that can absorb individual differences is used.

FIG. 2 is a diagram showing an example of an analysis method using a group of bandpass filters (B P F).
/) is analyzed using a group of 16-channel band-pass filters (all bands are 200 to 6000H2) (B P F
It is a time change diagram of the analyzed spectrum pattern.

One unit of the time axis is 18 m5, and if you take a cross section at a certain time, it becomes the spectrum at that time.Actual recognition processing is all digital processing, and at a certain time i
The intensity values of the spectrum in a horizontal row are expressed as a feature vector ai
(=ai1ai2 aI3- ai8″°°ai+6)
Then, the input voice pattern (here, the voice pattern of "3") is A=a1 a2-aj-aI (I=32).

Therefore, the voice pattern can be expressed as follows.

A=a1 a2-ai-aI (1) ai is a quantity representing the feature of the voice at time i, and is generally a vector value, and A is this feature vector a i (i = 1
−32 (when I = 32)], and ■
corresponds to the length of voice pattern A.

Also, Veg1-ru ai is called a feature vector, and a i
= (ai4, aI2-aiq-aid) =(
2). Q is the order of the vector, which in the example of FIG. 2 corresponds to the number of channels of the bandpass filter group, 16.

Similarly, let Bn be the standard pattern for word n.

Bn=bl' b2 n・=bjn-bJn'-
It is expressed as (3). At this time, bJn is a feature vector of the standard pattern of word n at time j, and has the same degree as the feature vector ai of the input cover turn A. Also, Jn represents the length of the standard pattern of word n, n is a serial number indicating the word name, and assuming a recognition word set of N words and Σ, Σ=(nln=1.2-N)... −・(4).

However, if there is no need to specify a specific word, omit the subscript n and write B=b, b2...bj...bJ (5
) t) J= (bji r t) J21...bj8.
...b jQ) ...(6).

In the speech recognition process, standard patterns Bn of all words in the recognized word set are time-normalized and pattern matched for the input cover turn A, and a word n closest to the input cover turn A is searched out of N words.

Figure 3 shows a mapping model for time normalization, which is
In the above example, the standard pattern B of the word "3" is aligned with the time axis of the input pattern using a mapping function,
Usually, the mapping function is expressed as j=j(i)...(
7) This is called the distortion function.

If this distortion function is known, the time axis of the standard pattern B is converted by equation (7) to calculate the time axis i of the input pattern A.
However, in reality, this distortion function is unknown, so one pattern is artificially distorted to be most similar to the other pattern, i.e. the distance is minimized to obtain the optimal distortion. I am trying to define a function.

FIG. 4 is a diagram for explaining an example of the DP matching method for carrying out the above principle. If we now consider the distortion function j(i) as a function of the time axis of standard pattern B without distortion, this distortion Pattern B is converted into the following pattern B' by function j(i).

B' = bj (+) bj (z)...bj(nil...b j (I)...(8) Above, the child distortion function is given by considering the time-only development of the actual speech pattern. For example, (i), j (i) is (approximately) a monotonically increasing function.

(b), j(1) is (approximately) a continuous function.

(c), j(1) takes a value near 1.

There are almost infinite number of distortion functions that satisfy these conditions, but among them, the distortion function j (i) for which B' is most similar to input pattern A, that is, the distance is the smallest. Establish. To do this, first, the time axis of standard pattern B is mapped onto the i-axis of input cover pattern A using distortion function j(1) to obtain pattern B', but at this time,
The optimal distortion function is the distortion function j (i) that minimizes the distance between pattern A and pattern B'. The distance between the input cover pattern A and the mapping pattern B' is Σ If ai −bj(i) If −−(9)1=
It is expressed as 1. Here, II II indicates the distance between the two vectors. The distance minimization problem in equation (9) above is defined as j=j(i) i=1 q=1. Generally, D (A, B) is called a time-normalized distance or inter-pattern distance, and d (i+J) is the distance between vectors ai and bj, which is usually called inter-vector distance.

Figure 5 abstracts the (i, j) plane shown in Figure 4 into a lattice plane, and calculates the coordinates (i, j) of each lattice point.
The vector quantity ratio W&d(i.

j), and if we consider the above equation (10) on this plane, we can start from (1, 1) and search for the optimal route (path) to (I, J). However, in this case, the path from state i-1 to state i is often limited to three as shown in the figure. Note that the matching window is used to prevent extreme time distortion, and the above three conditions (a) regarding time normalization are met when the matching window is used.
Satisfy ~(c) and ~j. Here, now i=1゜2
...Considering the case where the evaluation function of equation (10) is minimized by optimally controlling which state j to move to for each i in ■, the initial condition is g (1 ,1)=d(1,1)...(12)
The recurrence formula is D (A, B) = g (I, J) (14
), and the calculation of the previous equation 1sH3) is performed by tracing the lattice points in FIG. 5 as (i, j) increases. In other words, g(i, j) is the minimum sum of distances from point (1, 1) to point (i, j), and equation (13) is
j of the (i-1)th stage.

g that has already been determined for (j-1) and (j-2)
(l 1.j)+g(i 1y j 1)+g(11+
j 2), g (i+
j).

FIG. 6 is a block diagram of a processor that executes the above-mentioned DP matching process, in which 11 is a Δ memory, 12 is a B
memory, 13 is d(1+J) calculation unit, 14 is g(l,
j) The calculation unit 415 is a G(j) memory, 16 is a control unit,
d N+ j) The calculation unit 13 calculates the distance between the vectors al and bi, and the g (jrj) calculation unit 14 calculates the shortest distance g (i, j) to (i, j), and these are processed in parallel. do. g(i,''); When calculating j=I~J, use G (
j) In the memory 15, g (] 1+ j); J = 1~
It has a J in it. Moreover, min detects the smaller of gl and g2, and puts the smaller value into g.

Therefore, when using the above DP matching method, No. (13)
As is clear from the first term of the equation, if no matching window is provided, at least IXJXN calculations (where N is the number of registered words) are required.

In order to reduce the distance calculation amount by the above DP method.

A split method that takes onomatopoeic units has been proposed, but the spirit method calculates the distance between each frame of input speech in advance by calculating the distance between each frame of the input speech and a finite number of onomatopoeias (codebook). This method reduces the amount of distance calculation by simply searching the matrix for DP matching. In this split method, vector quantization is performed only on word standard patterns, and vector quantization is not applied to input speech. Therefore, in this Spirit method, a distance 71-ricks is created between the analysis frame of the input speech and the onomatopoeia (vector) stored in advance, but in this distance matrix, the horizontal axis is the frame number of the input speech. , the vertical axis is an onomatopoeic (vector) number, and by referring to this distance matrix, DP matching is performed between the standard pattern stored as a vector number series and the input speech.

A double split method has been proposed as a further improvement of the above split method, and this double split method is a method of vector quantizing input speech that includes only a standard pattern.

FIG. 7 is a block diagram for explaining an example of the double blit method, in which 20 is an input section, 21 is an analysis section, 22 is a vector quantization section, 23 is a vector number generation section, and 24 is a block diagram for explaining an example of the double blit method. 25 is a vector distance matrix table; 26 is a DP matching unit; the input audio is converted into a standard feature vector in the standard vector storage unit 24 and quantized in the vector quantization unit 22;
It is converted into a vector number sequence in the vector number generation section 23 and sent to the DP matching section 26. The DP matching unit 2G performs a D matching between the vector number sequence of the input voice sent as described above and the vector number sequence of the word standard button stored in advance in the vector storage unit 24.
P matching is executed while referring to the inter-vector distance matrix table 25 representing the distance between vectors.

To explain in more detail, the feature vector ai of the i-th frame of input audio A is a, i = (ai+, ai2, -aip) where i = 1.2. ... I (I; number of input frames P; number of dimensions of feature parameters; on the other hand, the second feature vector bk of the standard vector pattern B of speech is expressed as bk = (bk,,bk2...bkp ) However, k=
1.2. ... is represented by K (K: number of quantized standard vectors) (this characteristic vector bk has been vector quantized).

Second, in order to vector quantize the feature vector ai of the input voice, it is compared with the standard pattern in the standard vector storage section and converted into a feature vector bg.

However, i==arg m1nI(ai bk)l. However, i=arg rninl(ai-bk)1 refers to the value of k that makes l, :Li-bkl the most t, l\. That is, calculate the distance between ai and all bk, use the feature vector b↑ whose distance is closest to a] instead of 81, and '
Ai=b'i. Here, bk and bj (
k≠j) is calculated in advance and stored in the vector distance matrix table 25, the interframe distance between the feature vector ai of each input frame and the vector quantized word standard pattern bk is calculated in the table. 25. Here, the standard pattern of registered words is obtained by the vector number sequence of the efA quasi-vector.

The above double sprint method is (1) The amount of distance calculation can be reduced compared to the split method.

(2) Since the distance matrix can be set in advance, the distance matrix can be set skillfully.

(3) LPG by using a simple scale for input speech
Calculation of feature parameters in analysis can be omitted.

It has the following advantages, and is particularly useful for specific speaker recognition. However, even feature vectors corresponding to the same phoneme in the same word vary greatly depending on the speaker. Therefore,
Speech recognition for unspecified speakers often uses a plurality of standard patterns (multi-template method, KNN method, etc.).

However, the amount of calculation and memory increases in proportion to the number of standard patterns, which poses a practical problem.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, this method was developed for the purpose of improving recognition accuracy and recognition speed in a speech recognition device using the double split method.

1-11 The structure of the present invention will be described below based on embodiments.

FIG. 8 is a block diagram for explaining one embodiment of the speech recognition device according to the present invention. In the figure, 27 is a vector number sea fence and a class number sea fence section, and 28 is a vector distance matrix classified by class. Parts of the table that function similarly to those in Figure 7 are given the same reference numbers as in Figure 7. As shown in FIG. 9, the standard pattern consists of a vector number sea fence and a class number sea fence in which each frame is classified into classes. The inter-vector distance matrix table 28 is
As shown in Figure 10, there is a table with M classes, and element d of the matrix of class C is the onomatopoeic rhyme i when the input vector is the onomatopoeic j rhyme i, and the vector of the standard pattern of the class M is the onomatopoeic rhyme j. It represents the distance between vectors from when . DP
In the matching unit 26, the distance matrix table of the class corresponding to the standard pattern is selected from 28 using the class number sea fence 27, and DP matching with the input cover pattern is performed by referring to the vector number sequence of the pattern.

Effects Therefore, according to the present invention, input speech is encoded by a vector number sequence, the vector distance matrix table is classified in advance according to the type of sound, and the classified index (class number sequence) is used. Since the I)P matching is performed by selecting a desired matrix table based on the (given by the fence), speech recognition can be performed more quickly and accurately.

[Brief explanation of the drawing]

Figure 1 is a basic configuration diagram of the speech recognition device, Figure 2 is a diagram showing an example of speech analysis, Figure 3 is a mapping model for time normalization, and Figure 4 is time normalization using a distortion function. Figure 5
6 is a block diagram of a processor that performs DP matching processing; FIG. 7 is a block diagram for explaining an example of the double split method; FIG. 8 is a diagram for explaining one embodiment of the speech recognition device according to the present invention, FIG. 9 is a diagram showing an example of the configuration of a standard pattern used in implementing the present invention, and FIG. It is a figure which shows an example of the distance matrix table between vectors used for implementation. 20... Input unit, 21... Analysis unit, 22... Vector quantization unit, 23... Template (vector) unit, 24... Vector sea fence unit, 25... Vector distance matrix Table, 26... DP matching section, 27... Vector number sea fence and class number sea fence section, 28... Distance matrix table between vectors classified into classes. Figure l Figure 2 Figure 3 Yukafu Ig Turn B Figure 4 Figure 9 Figure 10 Procedural Amendment (Issue June 27, 1980 1, Incident Indication 1982 Patent Application No. 93571) 2. Name of the invention Speech recognition device 3. Relationship with the case of the person making the amendment Patent applicant Ota Ri Nakamagome Address 1-3-6 Nakamagome, Ota-ku, Tokyo Name (Name>
(674) Ricoh Co., Ltd. Representative Hama 1) Hiroshi (and 1 other person) 4. Agent Address 7, 807, Shatrain Yokohama, 1-2-7 Furo-cho, Naka-ku, Yokohama 231, Contents of amendment (1), The claims of the specification are amended as shown in the attached sheet. (2) "Following in an increasing direction" written on page 8, line 7 of the specification is amended to "tracing in an increasing direction." (3), page 9, lines 7 to 8, “(1)
3) Correct "from the first term of equation (13)" to "from the first term on the right side of equation (13)." (4) In the 19th line of page 10, change "26 is the DP matching unit," to "26 is the DP matching unit, 29 is the minimum distance word identification unit, 30 is the recognition result output unit," to correct. (5) "Execute while referring to" described in the 9th line of page 11 of the same page is changed to "Execute while referring to the word, the word with the minimum distance is determined by the word identification unit 29, and the recognition result output unit 30 Correct to "Output." (6), rlai-bkl described on page 12, line 7 of the same
The value of k that minimizes ” rl(ai-bk)l
The value of k that minimizes ”. (7), page 14, lines 10 to 11, “
As shown in FIG. 9, the vector number sea fence is corrected to ``as shown in FIG. 9, the vector number sea fence b- of word n''. ] (8), "consists of class numbered sea fences" written in lines 12 to 13 of the same
It consists of a class number sea fence C of classified words n. ”. (9), r class C described in lines 15 to 16 of the same
The element d of the matrix of is the input vector J, and the element d of the 71 helix of class m is the input vector J. (10), "Class M" written in the 17th line is corrected to "Class m". (11), FD as stated in page 15, lines 3 to 4
P matching is performed. " is subjected to FDP matching, and the word with the minimum distance is determined by the word identification unit 29,
The recognition result output section 30 outputs the result. ”. (12) The "type of sound" described in the 8th line to the 9th line is corrected to "type of phoneme". (13), "28...class classification table" written in lines 13 to 14 on page 16 of the same page is changed to "28...
・Classified vector distance matrix table, 29. Word identification unit, 30. Recognition result output unit. ”. (14) Figures 5, 7, 8, and 10 of the drawings will be corrected as shown in the attached sheet. Claims: Speech recognition characterized by a quantized standard pattern, quantization means for quantizing input speech, means for indexing quantized input speech, and a class/keyed vector. Device. Figure 51 Figure 10 Figure 7

Claims (1)

[Claims] It has a quantized standard pattern, a quantization means for quantizing the input audio, a means for converting the quantized input audio into an intex, and a vector distance matrix table. A speech recognition device characterized by performing verification using the index.