JPH0764590A - Voice recognizer - Google Patents

Abstract

(57) [Abstract] [Purpose] To improve the efficiency of probability calculation in a speech recognition system using a discrete hidden Markov model. [Structure] A plurality of discrete Hidden Markov Models having parameters corresponding to a plurality of states are prepared for each word (or syllable) to be recognized, and a vector quantization code time series obtained based on an input speech is prepared. And a plurality of discrete Hidden Markov Models for calculating probability and obtaining a recognition result based on the calculated probability value. The parameters of the discrete Hidden Markov Models are
Memory 5 (62) for collectively storing a series of addresses for each parameter relating to the same vector quantization code
The arithmetic processor 4 reads the parameters corresponding to each vector quantization code from the memory 5 in the order of addresses, and performs the probability calculation on the vector quantization code time series.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device,
In particular, the present invention relates to a voice recognition device suitable for high-speed voice recognition processing of large vocabulary using existing hardware.

[0002]

2. Description of the Related Art Generally, speech recognition requires a very large amount of processing, and it is required to improve the processing efficiency in order to realize a speech recognition apparatus. In particular, the Hidden Markov Model (discrete Hidden Markov Model), which has been widely used recently,
In the following, referred to as HMM), the model of each word or each syllable has a huge amount of parameters, and when recognizing the input speech, it is necessary to calculate the probability for all words (or syllables). Need to access. Also, in the probability calculation, it is necessary to calculate for each state of each model, and the amount of processing becomes enormous.

For details of the hidden Markov model, see, for example, "An Introduction to Hidden Markov Mode.
ls ", IEEE ASSP MAGAZINE, January 1
986, pp4-16.

Conventionally, as a method of reducing the amount of processing in speech recognition, a beam search method that terminates the calculation processing for a candidate that has been calculated halfway and considered to have a low possibility, and a method with a small amount of calculation are used in advance to recognize the object to be recognized. Preliminary selection methods have been attempted, in which a promising candidate is selected from all candidates and recognition processing is performed only on the selected candidate.

As an example of the beam search method, the Institute of Electronics, Information and Communication Engineers, D, Vol. J71-D No. 9
pp. 1650-1659, (1988-9) "High-speed DP matching by combining frame synchronization, beam search, and vector quantization", IEICE Transactions, D-2, Vol. J72-D-2 No. 8p
p. 1248-1255, (1989-8) "Discussion of threshold function of DP beam search". The basic idea of reducing the amount of processing is to cut off the calculation processing for candidates that have been calculated halfway and considered to be unlikely, and reduce the number of calculation targets. Although the examples described in these documents are targeted at DP matching as a recognition method, the reduction method itself can also be applied to HMMs.

On the other hand, as an example of the preselection method, a collection of proceedings of the Acoustical Society of Japan, 1-3-17, (1986-10).
There is a method as described in "Preliminary selection method using spectral dynamics for large vocabulary word speech recognition". In the above conventional example, a vector quantization codebook is prepared in advance for each word to be recognized, and after the end of the input speech is detected, the entire input speech is vector-quantized using each of the above codebooks. , The quantization distortion at this time is accumulated for each codebook, and only the accumulated values smaller than a certain threshold value are compared.

[0007]

Both of the above-mentioned prior arts find a part having a low possibility in a recognition object based on a partial processing result or a rough calculation result, and omit the processing of the part so that the whole processing is performed. It reduces the amount of processing. Of course, both of these methods are effective in reducing the amount of processing, but HMM
It does not attempt to improve the efficiency of the probability calculation itself in. Further, there is no mention of speeding up the processing speed from the viewpoint of the hardware configuration.

The object of the present invention is to improve the efficiency of the probability calculation itself in the HMM, to speed up the processing considered from the viewpoint of the hardware configuration, and to use existing hardware to make it compact and inexpensive. It is to realize a large vocabulary voice recognition device.

[0009]

In order to achieve the above object of the present invention, a speech recognition apparatus according to the present invention uses a discrete Hidden Markov Model having parameters corresponding to a plurality of states as a recognition target word (or For each syllabic), a probability calculation is performed on the vector quantization code time series obtained based on the input speech using the plurality of discrete hidden Markov models, and the calculated probability value is calculated. A speech recognition apparatus for obtaining a recognition result based on a calculation processor for performing the probability calculation, and parameters of the plurality of discrete Hidden Markov Models to a series of addresses for each parameter related to the same vector quantization code. And a memory for collectively storing the arithmetic processor,
With respect to the vector quantization code time series, the parameters corresponding to each vector quantization code are read from the memory in the order of addresses, and the probability operation is performed.

[0010]

In the speech recognition apparatus of the present invention, in executing the probability calculation using the HMM, the parameters (data) are arranged in the memory so that the data of the HMM can be efficiently accessed, and consecutive addresses are set. Since the memory that can be efficiently accessed is used, high-speed HMM probability calculation can be executed.

Further, the multiplication calculation of the transition probabilities and the appearance probabilities that can be calculated in advance due to the structure of the HMM (addition calculation if performed in the logarithmic domain) is performed in advance, and the result is stored in the parameters of the model. Every time, the probability is calculated using this at the time of recognition, so that the amount of calculation at the time of recognition can be reduced and the speed can be further increased.

Further, the probability of the corresponding HMM is judged using the appearance frequency information of the vector quantization code for each HMM, and the probability calculation is omitted for the HMM with a low probability, so that the speed is further increased. Can be achieved.

Furthermore, a large vocabulary can be dealt with by sharing words (or syllables) to be recognized by a plurality of arithmetic processors.

As described above, according to the present invention in which various speed-up techniques are integrated, it is possible to realize a compact, inexpensive and high-speed large-vocabulary speech recognition device by using the latest inexpensive hardware.

[0015]

EXAMPLES Examples of the present invention will be described below. The present invention can be applied to various types of speech recognition such as single syllable recognition, word recognition, and sentence recognition, but here, for the sake of simplicity, word recognition will be taken up and described.

FIG. 1 is a block diagram showing the hardware structure of the speech recognition apparatus of the present invention. The audio input from the microphone 1 is amplified by the audio amplifier 2. The amplified audio signal is taken in the AD converter 3 at regular time intervals (for example, 125 μs in 8 kHz sampling) and digitized. The digitized voice signal is subjected to various kinds of processing in the arithmetic processor unit 4 while referring to the contents of the external memory 5, and finally a recognition result is obtained.

The processing in the arithmetic processor unit 4 will be described with reference to the flowchart of FIG.

For simplification of the description, voice recognition of ten numbers from "0" to "9" will be described as an example.

A numerical voice, for example, "1" is input into the microphone.
("), A voice signal is taken in the AD converter 3 at regular time intervals (for example, every 83.3 μs in the case of 12 kHz sampling) and digitized (S31, S32). In the arithmetic processor unit 4,
The autocorrelation function is calculated every time sample data obtained by digitizing voice data is obtained (S33). The calculation of the autocorrelation function is performed partially every time one sample data is obtained. Expressed as an expression, it is as follows. Here, ri represents a variable for storing a partial result of the i-th-order autocorrelation coefficient, and xt represents sample data at time t. Note that the order of autocorrelation is 14th here.

Ri = ri + xt × xt-i For a predetermined number of data points (for example, when the analysis window length is 20 ms and 12 kHz sampling is performed, 2
When the above calculation is performed for 40 points), the autocorrelation function ri (i = 0 to 14) for one frame is determined, and the process proceeds in frame units (S35). If the period of the frequency analysis (frame period) is 20 ms, which is the same as the analysis window length, 20 m
The autocorrelation coefficient is determined once every s, and the processing is performed in frame units.

In the frame unit processing, first, a linear prediction coefficient is calculated from the autocorrelation function (S351), and a cepstrum coefficient is obtained from the linear prediction coefficient (S35).
2). The obtained cepstrum coefficient is regarded as a multidimensional vector, and vector quantization is performed using a vector quantization codebook prepared in advance to obtain a vector quantization code (S353). The vector quantization level (codebook size) can take any value, but is 256 in this embodiment. That is, after vector quantization, the quantization code k (any positive value from 1 to 256) is obtained. That is, assuming that the voice length of the input word voice "1" is L frames, a code sequence of length L is obtained.

In word speech recognition, the probability of outputting the above-mentioned code sequence of length L is calculated for discrete Hidden Markov Models (HMMs) of all recognition targets (all 10 numbers in this example) prepared in advance. The most probable HMM
Is the recognition result.

The actual HMM probability calculation process is not performed after the length L code sequence is obtained, but is performed each time one vector quantization code k is obtained, as shown in the flowchart of FIG. It is determined whether or not the voice end is detected (S36), and if detected, sorting /
Proceed to the candidate output process. While the voice end is not detected and the input voice continues, the flow returns to the beginning of the flowchart to continue the autocorrelation calculation and the frame unit processing.

When the above processing is performed and the end of the voice is detected, the probability calculation ends, and the probability value of each word (10 numbers from "0" to "9") is set as the score of each word. Each word is sorted based on (S37). The sorted upper L (eg L = 3) candidates are output as the recognition result (S38). For example, in the present example, if the probability value for the word “1” becomes high and the HMM is output as the first place, the correct answer is recognized.

For the processing for obtaining the linear prediction coefficient and the processing for obtaining the cepstrum coefficient, for example, the method described in Furui "Digital Speech Processing" Tokai University Press, etc. may be used. Further, in the flowchart of FIG. 3 of the present embodiment, the probability calculation using the HMM is configured to be performed in synchronization with other frame unit processing.
It is of course possible to make the probability calculation part using MM independent as another process and execute it in multiple processes.

Next, the HM in the flowchart of FIG.
The probability calculation part (S354) using M will be described in detail. First, the HMM will be described with reference to the HMM explanatory diagram of FIG. HMM has several states (number of states N
And ), The probability that the state transition will occur for each state transition (transition probability), and the probability that each vector quantization code will appear (occurrence probability) when the state transition occurs It is defined. Number of states N
Is, for example, about 20 for words and about 5 for syllables. HMM is a model that expresses speech, but when modeling with words as units (word HMM)
And a model for each small unit such as a syllable (syllable HM
M) A word may be represented by the combination of these small models. In this embodiment, the case of the word HMM will be considered. If the vocabulary to be recognized is M (for example, M = 10), M HMs
Prepare M.

Although the HMM is a state transition model having N states, the number N of states will be described as 5 here. Figure 4
Shown in (1) is one HMM having five states, which corresponds to a certain word w (any one of 1 to M). The state is indicated by a circle in FIG. 4, and the number in the circle corresponds to the state number.
Portions that allow transitions between states are connected by arrows (arcs). Generally, the HMM allows a state transition from an arbitrary state to an arbitrary state, but here, a left-to-right model often used in speech recognition is taken up. The left-to-right model allows only the state transition to itself and the state transition to the next state (state with a larger number). Each state transition is accompanied by a state transition probability (indicated by symbol a in the figure) and an appearance probability of each vector quantization code (indicated by symbol b in the figure) at that time.

Vector quantization code time series k (1), k
(2), k (3) ... k (t) (k (t) is 1 to 25)
6) and the HMM state i of the word w
Let us denote the probability of being at P (w, i, t). M
The problem of word voice recognition is P (w, N, T) (w = 1 to
It can be considered as a problem in which M) finds w that gives the maximum value. Therefore, the calculation of P (w, i, t) directly leads to the voice recognition process. There are several methods for calculating P (w, i, t), but here we will use a method called the Viterbi algorithm. Prior to the calculation, the initialization is performed according to the following formula.

P (w, i, 0) = 1 (i = 1, w = 1 to M) (1) P (w, i, 0) = 0 (i ≠ 1, w = 1 to M) ) (2) After that, vector quantization code k (t) (k (t) = 1 to
256) is obtained, the probability value is updated for each state of each word according to the following equation.

Wk1 = P (w, i-1, t-1) × a (w, i-1, i) × b (w, i-1, i, k (t)) (3) wk2 = P (w, i, t-1) × a (w, i, i) × b (w, i, i, k (t)) (4) P (w, i, t) = max (wk1, wk2) (5) where a (w, i, j) is from state i to state j of word w
, B (w, i, j, k (t)) is the probability that the vector quantized code k (t) appears in the transition from the state i to the state j of the word w. The above calculation flow is shown in a flowchart of FIG. If all probability values are expressed in the logarithmic domain, the above equation (3)
All multiplications during the probability calculation of (4) can be replaced with additions.

The above is the HMM probability calculation for one frame, and this is repeated until the voice end is detected, and finally P (w, N, T) is obtained for all M words, and the higher order of these is obtained. The recognition result can be obtained by selecting one.

Next, the arrangement of the HMM data in the memory will be described.

As shown in the description of the probability calculation using the HMM, the state transition probability a (w, i, j) (i = 1 to 5, j = i, i + 1) per HMM of one word.
Is 10 words, vector quantization code appearance probability b (w,
i, j, k) (i = 1 to 5, j = i, i + 1, k = 1 to
256) consists of 2560 words of data, totaling 2570 words. Various varieties are conceivable as to how this data is arranged in the memory.

In the simplest case, as shown in FIG. 6A, the words are first stored together and stored in each word, and each word is grouped into each state, and in each state, transition to the own state and the next state are stored. It is conceivable that the method is divided into two parts, that is, the transition probability a is first stored in each part, and subsequently, the probability b of appearance of the vector quantization code of 256 words is stored in the order of address. However, if the data is arranged in this way, it is necessary to access the discrete addresses in the memory in the probability update calculation of the above formulas (3) and (4), which is not efficient.

Therefore, the HMM data is rearranged to a large extent, as shown in FIG. 6B). That is, the HMM data is not sorted for each word, but for each vector quantization code. All HMM information about a specific vector quantization code k is stored collectively in a local address area. With the arrangement shown in b) of FIG. 6, when the vector quantization code k is determined, the HMM data necessary for the probability calculation in that frame are put together in a local area, and the equation (3 ) Since the data is stored in the address order in a form that matches the calculation order of (4), the reference of the HMM data during the probability calculation is executed only by incrementing the address. In the arrangement shown in b) of FIG. 6, the state transition probability a (w, i, j) is stored as a pair with the appearance probability b (w, i, j, k) of each vector quantization code. Although it is almost twice as large as the case of the storage method shown in 6 a), the calculation efficiency is high.

In FIG. 6 b), the HMM data is rearranged in the memory in advance so as to improve the calculation efficiency. However, as shown in FIG. 2, the data transfer control unit 63 is provided inside the arithmetic processor 6. , The data transfer control unit 63 fetches only the HMM data relating to the vector quantization code k from the external memory 5 every time the vector quantization result k is obtained, and outputs this to a series of addresses in the internal memory 62 of the arithmetic processor unit 61. Internal memory 62
If the HMM probability calculation is performed using, the same calculation efficiency can be achieved.

In the embodiment, as the internal memory or the external memory, a memory capable of efficiently accessing continuous addresses such as a RAMbus or a synchronous DRAM may be used. For such a memory, for example, Nikkei Electronics 1992.3.
16 (no549), pp. 95-97, Nikkei Electronics 1992.5.11 (no553), 143-14.
It is disclosed on page 7.

Next, in the probability calculation using the HMM, the pre-execution of the calculation which can be pre-executed due to the structure of the HMM calculation will be described. Looking at equations (3) and (4), array elements a (w, i, j) and b (w, which appear in the same equation)
It can be seen that the subscripts of i, j, k) are all the same except for k. That is, in the formulas (3) and (4), the multiplication of a (w, i, j) × b (w, i, j, k) ... (6) has a character that can be executed in advance, Formula (6)
Is calculated in advance, and the result is b ′ (w, i, j,
k), b ′ (w, i, j, k) = a (w, i, j) × b (w, i, j, k) (7) is stored as a new parameter, and Using HMM
The probability of can be calculated. This makes it possible to reduce the amount of calculation by almost half, as is clear from the comparison between the equations (3) and (4) and the equations (8) and (9). In probability calculation using HMM, the following equation (8) is used instead of equations (3) and (4)).
(9) will be used.

Wk1 = P (w, i-1, t-1) × b ′ (w, i-1, i, k (t)) (8) wk2 = P (w, i, t- 1) × b ′ (w, i, i, k (t)) ・ ・ ・ (9) P (w, i, t) = max (wk1, wk2) ・ ・ ・ (10) The arrangement of HMM data is as shown in FIG. That is, the transition probability data and the appearance probability data are preliminarily multiplied, and the information that conventionally requires two words is stored in one word. Therefore, the memory amount is also halved as compared with the arrangement in the case of FIG.

Next, an embodiment in which the number of words to be recognized is increased by using a plurality of arithmetic processors will be described with reference to FIG.

With a single arithmetic processor, the processing capacity is limited and the number of vocabularies that can be recognized is naturally limited. One solution is to use multiple arithmetic processors. Although the number of arithmetic processors is three in the embodiment of FIG. 8, the number of arithmetic processors is not particularly limited. In FIG.
The first arithmetic processor 71 performs almost the same processing as that of the arithmetic processor 4 in the embodiment shown in FIG. 1, but the vector quantization result is applied to all the other arithmetic processors 73 and 75.
The difference is that the HMM probability calculation result of the word for which another arithmetic processor is in charge is received from another arithmetic processor. The other arithmetic processors 73 and 75 receive the vector quantization result k from the first arithmetic processor 71 and use it to calculate the HMM probability. After the voice end is detected, the final result of the HMM probability calculation is returned to the first arithmetic processor 71. In the first arithmetic processor 71, the HMM probability calculation result of the word for which it is in charge and the H related to another word received from another arithmetic processor.
A recognition result is obtained by integrating all the MM probability calculation results. From the above, it is possible to easily deal with the increase in the number of vocabularies. In the case of the present embodiment, as shown in FIG. 8, each arithmetic processor is in charge of a certain number of words, but the HMM data of the in-charge word is stored in the external memory 72 provided individually for each arithmetic processor. 74 and 76. Therefore, in this case, it is necessary to store the data of the HMM for each word in charge in the external memories 72, 74, and 76, and the memory arrangement as shown in FIG. 6B) cannot be achieved. Considering both the efficiency of memory access and the independence of each word, the memory layout of the HMM data in this embodiment is as shown in FIG. That is, the HMM data is divided and held for each word, and is arranged for each vector quantization code within the word.

Next, HM whose appearance probability is lower than a certain standard
A method of reducing the overall calculation amount and speeding up the recognition processing by omitting the probability calculation for M will be described.

As can be seen from the equations (3) and (4), when the appearance probability for a particular vector quantization code has a very small value, the probability of the HMM becomes a very small value, and this HMM is the final value. It is less likely to remain as a recognition result. Therefore, the appearance probability of each vector quantization code in each HMM is checked in advance, and when a vector quantization result with a very low appearance probability is obtained, the HMM
The probability calculation of can be omitted. For each vector quantization code, HMMs having transitions whose appearance probability is lower than a predetermined reference are listed to create a table as shown in FIG. At the time of recognition, the table of FIG.
MM is obtained, and probability calculation is omitted for this HMM. As a result, the probability calculation can be largely omitted and faster voice recognition can be performed.

The flow of voice recognition when this method is introduced is shown in the flowchart of FIG. This flowchart is obtained by inserting steps S111 and S112 for omitting calculation based on the occurrence probability into the flowchart shown in FIG.

Next, a method of omitting the probability calculation of the HMM by using the statistical information of the vector quantization code in the section of constant time length will be described.

As a fixed time length, for example, as shown in FIG. 12A, 10 frames before and 10 frames after the frame to be processed, a total of 21 frames (about 4 when the frame period is 20 ms).
Consider the 00 ms interval). A histogram as shown in b) of FIG. 12 is calculated as statistical information. This histogram can be obtained by simply counting the number of each of the 21 quantized codes. On the other hand, each HMM has an appearance probability of each vector quantization code for each transition, which corresponds to the histogram as it is. Therefore, if the histogram for each transition is averaged over all transitions, a histogram can be obtained, and this can be considered as a histogram of this HMM.
The two histograms are compared, and when the similarity is lower than a predetermined criterion, the probability calculation for that HMM is omitted.

As a similarity measure, for example, a method of considering the histogram as a multidimensional vector and taking an inner product can be considered. If the similarity measure calculated in this way is smaller than a preset reference value (for example, 0.1), H
The calculation of the probability of MM is omitted.

From the above, the probability calculation can be largely omitted,
Higher speed voice recognition is possible. The flow of speech recognition when this method is introduced is shown in the flowchart of FIG. This flowchart is obtained by inserting steps S131 and S132, which omit the calculation using the time-series statistical information of the vector quantization code, into the flowchart shown in FIG.

[0049]

As described above, according to the present invention, in executing the probability calculation using the HMM, the data is arranged in the memory so that the data of the HMM can be accessed efficiently, and the continuous addresses are efficiently accessed. Since it is configured by using a memory that can be used, high-speed HMM probability calculation can be executed. In addition, the amount of calculation at the time of recognition can be reduced by performing all the calculations that can be performed in advance due to the structure of the HMM. Further, by using the appearance frequency information of the vector quantization code for each HMM, the probability calculation can be omitted for the HMMs with low possibility.

As described above, according to the present invention in which various speed-up techniques are integrated, it is possible to realize a compact, inexpensive and high-speed large-vocabulary speech recognition device by using the latest inexpensive hardware.

[Brief description of drawings]

FIG. 1 is a block diagram showing a hardware configuration of an embodiment of a voice recognition device of the present invention.

FIG. 2 is a block diagram showing a hardware configuration of another embodiment of the voice recognition device of the present invention.

FIG. 3 is a flowchart showing an outline flow of processing of one embodiment of the voice recognition device of the present invention.

FIG. 4 is an explanatory diagram illustrating a discrete hidden Markov model used in the speech recognition apparatus of the present invention.

FIG. 5 is a flowchart showing a detailed procedure of probability calculation processing by a discrete hidden Markov model used in the speech recognition apparatus of the present invention.

FIG. 6 is an explanatory view for explaining how the data of the discrete hidden Markov model used in the speech recognition apparatus of the present invention is arranged in the memory.

FIG. 7 is an explanatory diagram illustrating pre-calculation of state transition probabilities and appearance probabilities of discrete Hidden Markov Models used in the speech recognition apparatus of the present invention.

FIG. 8 is a block diagram illustrating an embodiment of a plurality of arithmetic processors of a voice recognition device of the present invention.

FIG. 9 is an explanatory view for explaining how the data of the discrete Hidden Markov Model is arranged in the memory in the embodiment by the plural arithmetic processors of the speech recognition apparatus of the present invention.

FIG. 10 is an explanatory diagram of a table listing discrete Hidden Markov Models with low occurrence probability of vector quantization code.

FIG. 11 is a flowchart illustrating a procedure for omitting a part of probability calculation by a discrete hidden Markov model based on the appearance probability of a vector quantization code.

FIG. 12 is an explanatory diagram illustrating a method of omitting a part of the probability calculation by the discrete hidden Markov model based on the statistical information of the vector quantization code time series.

FIG. 13 is a flowchart illustrating a procedure of omitting a part of probability calculation by a discrete hidden Markov model based on statistical information of vector quantization code time series.

[Explanation of symbols]

1 ... Microphone, 2 ... Audio amplifier, 3 ...
AD converter 4 ... Arithmetic processor, 5 ... External memory, 61 ...
..Arithmetic processor 62 ... internal memory, 63 ... data transfer control unit 71 ... first arithmetic processor, 72 ... first external memory 73 ... second arithmetic processor, 74 ... ..Second external memory 75 ... Third arithmetic processor, 76 ... Third external memory

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hiroshi Ikeda 1-280 Higashi Koikeku, Kokubunji, Tokyo Inside Central Research Laboratory, Hitachi, Ltd. (72) Toshiyuki Arizuka 1-280 Higashi Koikeku, Kokubunji, Tokyo Hitachi Ltd. (72) Inventor, Kazuyuki Kodama, Kazuyuki Kodama, 1-280, Higashi Koikeku, Kokubunji, Tokyo Hitachi, Ltd., Central Research Institute (72) Takaki Noguchi, 1-280, Higashi Koikeku, Kokubunji, Tokyo Hitachi, Ltd. Central Research Co., Ltd. In-house

Claims (11)

[Claims] 1. A vector-quantized code time series obtained based on an input speech by preparing a plurality of discrete Hidden Markov Models for each recognition target word (or syllable) having parameters corresponding to a plurality of states. A plurality of discrete Hidden Markov models are used for probability calculation, a speech recognition apparatus for obtaining a recognition result based on the calculated probability value, an arithmetic processor for performing the probability calculation, A plurality of discrete Hidden Markov Model parameters, a memory for collectively storing in a series of addresses for each parameter related to the same vector quantization code, the arithmetic processor, for the vector quantization code time series And read the parameters corresponding to each vector quantization code from the memory in the order of addresses and perform the probability operation. Speech recognition apparatus according to claim. 2. The voice recognition device according to claim 1, wherein the memory is an external memory of the arithmetic processor. 3. The voice recognition device according to claim 1, wherein the memory is an internal memory of the arithmetic processor. 4. An external memory that stores parameters of the plurality of discrete Hidden Markov Models, an internal memory that stores parameters that are targets of probability calculation by the arithmetic processor, and exists at distributed addresses of the external memory. 4. The voice recognition apparatus according to claim 3, further comprising a data transfer control unit that takes out a parameter corresponding to a specific vector quantization code and transfers the parameter to a series of addresses in the internal memory. 5. The discrete Hidden Markov Model has a state transition probability regarding a state transition in each state and an appearance probability of each vector quantization code in the state transition as the parameter. The voice recognition device described. 6. The state transition probability for each state transition of the discrete Hidden Markov Model for each word (or each syllable) and the appearance probability of each vector quantization code in the state transition are stored in the memory in advance. The speech recognition apparatus according to claim 5, wherein the multiplied value is stored as a new parameter in place of both parameters, and the probability calculation is performed using this parameter. 7. A table is prepared in which discrete Hidden Markov Models whose appearance probability of the vector quantization code is smaller than a predetermined value are arranged and stored for each vector quantization code, and the arithmetic processor is configured to: Each time a vector quantization code is obtained, the table is referred to, and the probability calculation is omitted for the discrete hidden Markov model existing in the corresponding vector quantization code column in the table. The voice recognition device according to claim 5 or 6. 8. The arithmetic processor calculates statistical information for a certain time length of the obtained vector quantization code time series, and the statistical information satisfies a predetermined criterion for each discrete hidden Markov model. The speech recognition apparatus according to claim 5 or 6, characterized in that the probability calculation for the model is omitted in the case. 9. A microphone for inputting voice, an audio amplifier for amplifying a voice signal input from the microphone, an AD converter for fetching and digitizing the amplified voice signal at regular time intervals, and a digital signal for fetching. What is claimed is: 1. A voice recognition device, comprising: an arithmetic processor for performing an arithmetic operation on a converted voice signal; and a memory for storing various data used by the arithmetic processor, wherein the arithmetic processor is the voice captured by the AD converter. The signal is subjected to short-time frequency analysis to obtain the time series of the frequency spectrum, and the time series of the frequency spectrum is subjected to vector quantization using a codebook stored in advance in the internal memory of the arithmetic processor. To obtain the vector quantization code time series, and further to the memory for the vector quantization code time series. For each word (or syllable) stored in the memory, probability calculation is performed using a discrete Hidden Markov Model, and the recognition result is obtained based on the calculated probability value. The speech recognition device characterized in that the parameters of the discrete Hidden Markov Model of (1) are stored together in a series of addresses for each parameter related to a specific vector quantization code. 10. A microphone for inputting a voice, an audio amplifier for amplifying a voice signal input from the microphone, an AD converter for fetching and digitizing the amplified voice signal at regular time intervals, and a digital signal for fetching. A plurality of arithmetic processors for performing arithmetic on the converted audio signal,
A voice recognition device comprising a plurality of memories for individually storing various data used by each arithmetic processor, wherein the first arithmetic processor converts a voice signal taken in by the AD converter. On the other hand, a short-time frequency analysis is performed to obtain the time series of the frequency spectrum, and the time series of the frequency spectrum is vector-quantized using a codebook stored in advance in the internal memory of the arithmetic processor to obtain the vector. Obtaining the quantization code time series, further performing probability calculation using the parameters of the discrete hidden Markov model for each word (or syllable) stored in advance in the memory for the vector quantization code time series, The obtained vector quantization code is sent to all other arithmetic processors, and all other arithmetic processors The processor receives the vector quantization code from the first arithmetic processor, and performs probability calculation using the parameters of the discrete hidden Markov model for each word (or syllable) stored in advance in the memory for each arithmetic processor. The probability value finally obtained is sent to the first arithmetic processor, and the first arithmetic processor combines all the probability values calculated by itself and the probability values received from other arithmetic processors. A voice recognition device characterized in that a recognition result is obtained. 11. Each of the memories stores parameters of a discrete hidden Markov model of a word (or syllable) which the corresponding arithmetic processor is in charge of, and at this time, a vector quantization code for each word (or syllable) is stored. 11. The voice recognition device according to claim 10, wherein the same parameters are collectively stored in a series of addresses.