JPH06175685A - Pattern recognition device and hidden markov model generating device - Google Patents

Abstract

PURPOSE:To simplify the constitution of the entire device and to reduce the number of arithmetic operations by beforehand computing weighting coefficient vectors and eliminating a weighting coefficient vector computing section. CONSTITUTION:The device is provided with a weighting coefficient vector storage means so as to store weighting coefficient vectors prior to the learning of a Hidden Markov Model(HMM) and to use their fixed values during a pattern recognition. Namely, a weighting vector storage section 104 stores weighting coefficients u1, u2,... UK corresponding to a first to a Kth representative vectors as fixed values independent to the frames of an input voice. Therefore, during an HMM learning, the computations of weighting coefficient vectors, which are computed sequentially, are eliminated while conducting a recognition, a voice recognition is performed by employing a smaller amount of a storage capacity, the device is simplified and the amount of computations is reduced.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition device and a Hidden Markov Model (HMM) creating device (hereinafter simply referred to as an HMM creating device), and more particularly to a device for identifying a time series pattern such as voice recognition. Regarding

The present invention can be applied to general time-series signals, but for convenience of explanation, the prior art and the present invention will be described below by taking speech recognition as an example.

[0003]

BACKGROUND OF THE INVENTION In general, speech recognizers convert an unknown speech signal into a sequence of defined acoustic feature vectors and then compare it with a previously stored acoustic feature vector representing an identified reference pattern. Is configured.
As a result of the comparison, the unknown speech signal is identified as the best matching reference pattern according to the defined recognition criteria. Currently, the reference pattern that is considered to have the best performance is H that uses a set of states and state transitions based on statistical estimation.
It is MM.

Therefore, first, the HMM will be described.
The HMM is used to evaluate an observation sequence such that each observation is one of a finite number M of symbols in the observation sequence O = o ₁ , o ₂ , ..., O _T. The observation sequence can be modeled as a stochastic function of a latent Markov chain with state transitions that are not directly observable. FIG. 5 is a diagram for explaining such an HMM.

In FIG. 5, the number of states N = 3 and the finite number of output symbols M = 4 are taken as an example. The transition between states 1, 2 and 3 is represented as a state transition probability matrix A = [a _ij ], where a _ij is the probability of transition to state j when the model is in state i. The probability of the output symbol of the model is represented by a symbol output probability matrix B = [b _j (k)], and b _j (k) is the probability of outputting the symbol k when the model transits to the state j. Create one HMM for each vocabulary,
It can be used to classify observation sequences based on the probability of generating an unknown observation sequence from each HMM.

A symbol sequence obtained for an unknown input is O = o ₁ , o ₂ , ..., O _T , and an arbitrary state sequence of length T that can be generated from HMMλ is S = s ₁ , s ₂ ,. _{When T} , the probability (likelihood) that the symbol sequence O occurs from λ is
It is shown as (Equation 1).

[0007]

[Equation 1]

In the above, the input x _t is converted into a unique symbol o _t , but it is converted into a plurality of K symbols o _t1 , o _t2 , ..., O _tK (the symbol vector q _t =
(o _t1 , o _t2 , ..., O _tK )), and a method of using the weight coefficients u _t1 , u _t2 , ..., U _tK with each symbol (weight coefficient vector v _t (u _t1 , u _t2 , ...). , U _tK )), in which case
The unknown input is a symbol vector sequence Q = q ₁ , q ₂ , ..., q
_T and the load coefficient vector sequence _{V = v 1, v 2,} ..., expressed in v _T, the probability of occurrence of the symbol vector sequence to the HMMramuda (likelihood) is expressed as shown in equation (2).

[0009]

[Equation 2]

Recognition is performed by comparing the likelihoods obtained from the HMMs of each vocabulary.

[0011]

[Outer 1]

[0012]

[Equation 3]

FIG. 6 is a block diagram showing the structure of a conventional speech recognition apparatus using an HMM. In the figure, reference numeral 601 denotes a feature extraction unit, which uses a well-known method such as linear predictive coding (LPC) analysis or Fourier transform of an input speech signal v to obtain a feature vector sequence X = x at regular time intervals. _{1, x 2, ... x t} , ..., is converted to x _T. Here, T is the length of the feature vector sequence in the input audio signal v.

Reference numeral 602 is called a codebook, and holds a representative vector representing a finite number M of symbols as shown in FIG. In other words, the symbol is stored in the first column of each row, and the representative vector is stored after that in the number M rows of the codebook.

Reference numeral 603 denotes a vector quantizer, which replaces the feature vector x _t with the symbols of the representative vectors of the 1st to Kth ranks in the closest order of the codebook 602, and symbol vectors q _t = (o _t1 , o _t2 , ..., o _tk ..., o _tK ), and the sequence of the feature vectors is converted into a symbol vector sequence Q = q ₁ , q ₂ , ..., q _T.

Reference numeral 604 denotes a weighting factor vector calculating unit, which calculates a weighting factor for each of the 1st to Kth representative vectors selected by the vector quantization unit 603 of the feature vector x _t according to (Equation 4), Coefficient vector v _t =
(u _t1 , u _t2 , ..., U _tk ..., U _tK ), and the load coefficient vector series V = v ₁ , v ₂ , ..., V _T is calculated.

[0017]

[Equation 4]

[0018]

[Outside 2]

[0019]

[Outside 3]

[0020]

[Equation 5]

[0021]

[Outside 4]

[0022]

[Equation 6]

Reference numeral 608 denotes a likelihood storage unit, which is the likelihood calculation unit 6
It is stored to compare the likelihood of each word calculated in 07.

Reference numeral 609 denotes a comparison / determination unit, which is the likelihood storage unit 6
The vocabulary corresponding to the HMM that gives the maximum likelihood value for each HMM stored in 08 is determined as the recognition result (rec).

The respective units 606 to 608 perform once for each HMM of each vocabulary and repeat from w = 1 to W, and the result is evaluated by the comparison and determination unit 609.

In order to perform recognition using the HMM as described above, it is necessary to create the HMM in advance. This is called HMM learning, and the method will be described below.

[0027]

[Outside 5]

Reference numeral 802 is called a codebook, which holds a representative vector representing a finite number of M symbols, and its configuration is the same as that shown in FIG.

[0029]

[Outside 6]

[0030]

[Outside 7]

[0031]

[Outside 8]

Reference numeral 806 denotes an HMM temporary storage unit, which is an initial HM.
M (where A and B are random numbers or empirical values) and learning H before the learning converges when repeating the learning H
The MM is stored, and the state transition probability matrix A and the symbol output probability matrix B are stored and updated every time learning is completed.

[0033]

[Outside 9]

[0034]

[Outside 10]

[0035]

[Equation 7]

[0036]

[Outside 11]

Reference numeral 810 denotes a re-estimation unit which calculates the state transition probability a _ij according to (Equation 8) and the symbol output probability b _{ij according} to ( _Equation 9).
_It re-estimates _i (m).

[0038]

[Equation 8]

[0039]

[Equation 9]

Reference numeral 811 is a learning convergence confirmation unit, which is a re-estimation unit 810.
It is determined from the state in 1) whether the learning is in the convergent state, and if it is in the convergent state, the convergent signal y is sent, and otherwise, the re-estimation command signal n is sent to the re-estimated HMM storage unit 812.

The re-estimation HMM storage unit 812 temporarily stores the re-estimated HMM, and the learning convergence confirmation unit 81
With the signal from 1, if the convergence signal is y, the re-estimation HMM is stored in the HMM storage unit 605 in FIG. 6, and if it is the re-estimation command signal n, it is stored in the HMM temporary storage unit 806.

The respective units 807 to 810 are repeated until the learning convergence confirmation unit 811 obtains the convergence signal y.

The above is the configuration of the speech recognition apparatus and the HMM creation apparatus using the conventional HMM.

[0044]

The weighting factor vector calculation units 604 and 804 in FIGS. 6 and 8 using the conventional speech recognition and the like as described above perform the calculation as shown in (Equation 4). Therefore, there is a problem that the configuration becomes complicated and the number of calculations increases.

An object of the present invention is to solve the problem by simplifying the weighting factor vector calculation unit by deleting the weighting factor vector in advance and simplifying the overall construction of the apparatus and reducing the number of calculations. .

[0046]

The invention according to claim 1 of the present invention has a weighting coefficient vector storage means, which stores a weighting coefficient vector in advance before learning of an HMM (Hidden Markov model) to recognize a pattern. The pattern recognition device is characterized in that the fixed value is sometimes used.

The invention according to claim 2 of the present invention further comprises a weighting coefficient vector storage means for storing the weighting coefficient vector in advance before learning the HMM (Hidden Markov model), and at the time of learning the Hidden Markov model. It is a Hidden Markov model creation device characterized by using the fixed value.

[0048]

According to the present invention, the weighting coefficient vector calculated before the HMM learning is stored in the weighting coefficient vector storage unit, and the calculation of the weighting coefficient vector which is sequentially calculated during the HMM learning and the recognition is deleted. However, the device can be simplified and the calculation amount can be reduced.

[0049]

EXAMPLES The present invention will be described below with reference to examples.

FIG. 1 shows H in the first embodiment of the present invention.
It is a block diagram which shows the structure of the speech recognition apparatus using MM. In the figure, 101 is a feature extraction unit, which uses a well-known method such as LPC analysis or Fourier transform of the input voice signal v to obtain a feature vector sequence X = x ₁ ,
_{x 2, ... x t, ...} , it is converted to x _T. Here, T is the length of the feature vector sequence in the input audio signal.

Reference numeral 102 is a codebook, which holds a representative vector representing a finite number of M symbols, and its configuration is the same as that shown in FIG.

Reference numeral 103 denotes a vector quantizer, which replaces the feature vector x _t with the symbols of the representative vectors of the 1st to Kth ranks in the closest order of the codebook 102, and symbol vectors q _t = (o _t1 , o _t2 , ..., o _tk ..., o _tK ), and the sequence of the feature vectors is converted into a symbol vector sequence Q = q ₁ , q ₂ , ..., q _T.

Reference numeral 104 denotes a weighting factor vector storage unit, which is a feature of the present invention, and weighting factors (u ₁ , ..., U _k ..., U _K ) for each of the _1st to Kth representative vectors are input to the frame of the input speech. It is stored as an irrelevant fixed value. The value may be determined by, for example, the reciprocal of the first place to the Kth place, or a value that gradually decreases from the first place to the Kth place may be appropriately given.

[0054]

[Outside 12]

[0055]

[Outside 13]

[0056]

[Equation 10]

[0057]

[Outside 14]

Reference numeral 108 denotes a likelihood storage unit, which is the likelihood calculation unit 1
It is stored to compare the likelihood of each word calculated in 07.

Reference numeral 109 denotes a comparison / determination unit, which is the likelihood storage unit 1
The vocabulary corresponding to the HMM that gives the maximum likelihood value for each HMM stored in 08 is determined as the recognition result (rec).

The respective units 106 to 108 perform once for each HMM of each vocabulary and repeat from w = 1 to W, and the result is evaluated by the comparison and determination unit 109.

As described above, in the present embodiment, at the time of voice recognition, the calculation of the weighting factor vector calculation unit, which has been sequentially calculated, can be deleted.

[0062]

[Outside 15]

Reference numeral 202 denotes a codebook, which holds a representative vector representing a finite number of M symbols, and its configuration is the same as that shown in FIG.

[0064]

[Outside 16]

Reference numeral 204 denotes a weighting coefficient vector storage unit, which is a feature of the present invention, in which the weighting coefficients (u ₁ , u ₂ , ..., U _k ..., U _K ) for each of the _1st to Kth representative vectors are input and learned. It is stored as a fixed value irrelevant to the frame of the audio signal v '. The method of determining this value may be, for example, the reverse of 1st to Kth, or a value that gradually decreases from the 1st to Kth may be appropriately given.

[0066]

[Outside 17]

Reference numeral 206 denotes an HMM temporary storage unit, which is an initial HM.
M (where A and B are random numbers or empirical values) and learning H before the learning converges when repeating the learning H
The MM is stored, and the state transition probability matrix A and the symbol output probability matrix B are stored and updated every time learning is completed.

[0068]

[Outside 18]

[0069]

[Outside 19]

[0070]

[Outside 20]

A re-estimation unit 210 re-estimates the state transition probability a _ij according to (Equation 8) and the symbol output probability b _i (m) according to ( _Equation 9).

Reference numeral 211 denotes a learning convergence confirming unit, which determines whether or not learning is in a convergent state. If the learning is in a convergent state, the convergent signal y is re-estimated.
It is sent to the MM storage unit 212.

The re-estimation HMM storage unit 212 temporarily stores the re-estimated HMM, and the learning convergence confirmation unit 21
With the signal from 1, if the convergence signal is y, the re-estimation HMM is stored in the HMM storage unit 105 in FIG. 1, and if it is the re-estimation command signal n, it is stored in the HMM temporary storage unit 206.

The respective units 207 to 210 are repeated until the learning convergence confirmation unit 211 obtains a convergence signal y.

The above is the HMM of each first embodiment of the present invention.
2 is a configuration of a voice recognition device and an HMM creation device using the.

As can be seen from the first embodiment described above, the conventional load coefficient vector calculation units 604 and 804 shown in FIG. 6 and FIG.
Is reduced, and instead the weighting factor vector storage unit 10
4,204 are given. The former is configured as a computing device, but the latter can store at most K values,
The structure is greatly simplified. In addition, it is possible to reduce the amount of calculation without having to perform calculation.

The experiment conducted by using the above-described embodiment of the present invention will be described.

As the vocabulary to be recognized, 100 place names of Japan were used, and a total of 54 words for each vocabulary uttered by 27 men twice were used as the learning data of each HMM. For this purpose, a total of 9600 words were spoken twice by 48 non-learning speakers. As for the result, as shown in (Table 1), the performance is not deteriorated as compared with the conventional method even though the calculation amount is reduced.

[0079]

[Table 1]

FIG. 3 shows H in the second embodiment of the present invention.
It is a block diagram which shows the structure of the speech recognition apparatus using MM. The configuration of the second embodiment is the same as the configuration of the first embodiment of FIG. 1, except that the weighting coefficient vector storage unit 305 stores the value of the weighting factor in advance in the HMM learning. It is a build of 304. In the figure, reference numeral 301 denotes a feature extraction unit that inputs the input voice signal v to the LPC.
A series of feature vectors X = x ₁ , x ₂ , ... _Xt , ..., At regular time intervals, by well-known methods such as analysis and Fourier transform.
Convert to x _T. Here, T is the length of the feature vector sequence in the input audio signal.

Reference numeral 302 is a codebook, which holds a representative vector representing a finite number of M symbols, and its configuration is the same as that shown in FIG.

Reference numeral 303 denotes a vector quantizer, which replaces the feature vector x _t with the symbols of the representative vectors from the 1st to the Kth in the closest order of the codebook 302, and symbol vectors q _t = (o _t1 , o _t2 , ..., o _tk ..., o _tK ), and the feature vector sequence is converted into a symbol vector sequence Q = q ₁ , q ₂ , ... q _t , q _T.

Reference numeral 304 denotes a pre-weighting coefficient vector calculation unit, which is a feature of the present invention, and calculates a weighting coefficient using the HMM learning data D and the codebook 302 according to (Equation 11).

[0084]

[Equation 11]

Reference numeral 305 denotes a weighting coefficient vector storage section, which is a weighting coefficient (u ₁ , ..., U _k) for each of the representative vectors of 1st to Kth ranks obtained in advance by the aforesaid weighting coefficient vector calculation section 304. , U _K ) is stored as a fixed value irrelevant to the frame of the input audio signal v.

[0086]

[Outside 21]

[0087]

[Outside 22]

[0088]

[Outside 23]

A likelihood storage unit 309 includes the likelihood calculation unit 3
It is stored in order to compare the likelihood of each word calculated in 08.

Reference numeral 310 denotes a comparison / determination unit, which is the likelihood storage unit 3
The vocabulary corresponding to the HMM that gives the maximum likelihood value for each HMM stored in 09 is determined as the recognition result (rec).

The respective units 307 to 309 perform once for each HMM of each vocabulary and repeat from w = 1 to W, and the result is evaluated by the comparison / determination unit 310.

[0092]

[Outside 24]

Reference numeral 402 denotes a codebook, which holds a representative vector representing a finite number of M symbols, and its configuration is the same as that shown in FIG.

[0094]

[Outside 25]

Reference numeral 404 denotes a pre-weighting coefficient vector calculation unit, which is a feature of the present invention, and calculates the weighting coefficient using the HMM learning data D and the codebook 402 in accordance with (Equation 11).

Reference numeral 405 denotes a weighting coefficient vector storage unit, which weights the weighting factors (u ₁ , u ₂₎ for the respective representative vectors of the 1st to Kth positions obtained in advance by the preloading coefficient vector calculation unit 304 of FIG. , ..., u _k ..., in which stored as no fixed value related to the frame of the u _K) input training speech signal v 'a.

[0097]

[Outside 26]

Reference numeral 407 denotes an HMM temporary storage unit which stores the initial HM.
M (where A and B are random numbers or empirical values) and learning H before the learning converges when repeating the learning H
The MM is stored, and the state transition probability matrix A and the symbol output probability matrix B are stored and updated every time learning is completed.

[0099]

[Outside 27]

[0100]

[Outside 28]

[0101]

[Outside 29]

A re-estimation unit 411 re-estimates the state transition probability a _ij according to (Equation 8) and the symbol output probability b _i (m) according to ( _Equation 9).

A learning convergence confirmation unit 412 determines whether or not the learning is in the convergent state. If the learning is in the convergent state, the convergent signal y is re-estimated.
It is sent to the MM storage unit 413.

The re-estimation HMM storage unit 413 temporarily stores the re-estimated HMM, and the learning convergence confirmation unit 41
In accordance with the signal from 2, if the convergence signal is y, the re-estimation HMM is stored in the HMM storage unit 306 in FIG. 3, and if it is the re-estimation command signal n, it is stored in the HMM temporary storage unit 407.

The respective units 408 to 411 are repeated until the learning convergence confirmation unit 412 obtains the convergence signal y.

The above is the configuration of the voice recognition apparatus using the HMM and the HMM creation apparatus of the second embodiment of the present invention.

As can be seen from the second embodiment described above, the conventional load coefficient vector calculation units 604 and 804 shown in FIG. 6 and FIG.
Is reduced, and instead, the weighting factor vector storage unit 30
5,405 have been given. The former is configured as a computing device, but the latter can store at most K values,
The structure is greatly simplified. In addition, it is possible to reduce the amount of calculation without having to perform calculation.

The experiment conducted using the present invention will be described.

As the recognition target vocabulary, 100 place names of Japan were used, and a total of 54 words for each vocabulary that 27 males uttered twice were used for the learning data of each HMM, and the recognition data was 100 vocabulary for each. For this purpose, a total of 9600 words were spoken twice by 48 non-learning speakers. As for the results, as shown in (Table 2), there is almost no deterioration in performance compared with the conventional method, even though the calculation amount is reduced.

[0110]

[Table 2]

[0111]

As described above, according to the present invention, by the calculation of the weighting coefficient vector in advance, the recognition rate is substantially unchanged as compared with the conventional one, and the pattern recognition is simpler than the conventional one and the number of calculations is small. It is possible to create a device.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a voice recognition device using an HMM according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of an HMM creating apparatus used for voice recognition in the first exemplary embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a voice recognition device using an HMM according to a second embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of an HMM creating apparatus used for voice recognition in a second exemplary embodiment of the present invention.

FIG. 5 is a diagram for explaining a Bidun Markov model (HMM).

FIG. 6 is a block diagram showing a configuration of a speech recognition apparatus using a conventional HMM.

FIG. 7 is a diagram showing a configuration example of a codebook.

FIG. 8 is a block diagram showing a configuration of a conventional HMM creating apparatus used for speech recognition.

[Explanation of symbols]

101, 201, 301, 401 ... Feature extraction unit, 102, 202, 302, 4
02 ... Codebook, 103, 203, 303, 403 ... Vector quantizer, 104, 204, 305, 405 ... Weighting coefficient vector memory, 106, 207, 307, 408 ... Weighted probability sum calculator,
105, 306 ... HMM storage unit, 107, 308 ... Likelihood calculation unit,
108, 309 ... Likelihood storage unit, 109, 310 ... Comparison determination unit, 20
5,406 ... HMM learning data storage unit, 206,407 HM
M temporary storage unit, 208, 409 ... Path probability calculation unit, 209, 4
10 ... Path probability storage unit, 210, 411 ... Re-estimation unit, 211, 4
12 ... Learning convergence confirmation unit, 212, 413 ... Re-estimation HMM storage unit, 304, 404 ... Pre-load coefficient vector calculation unit.

Claims (4)

[Claims] 1. A pattern recognition device comprising a weighting coefficient vector storage means, which stores a weighting coefficient vector in advance before learning a Hidden Markov model, and uses a fixed value when recognizing a pattern. 2. A Hidden Markov model is created, which has a weighting coefficient vector storage means, stores a weighting coefficient vector in advance before learning a Hidden Markov model, and uses a fixed value when learning the Hidden Markov model. apparatus. 3. A pre-load coefficient vector calculation means is provided,
The weighting factor vector storage means stores the weighting factor as an average of the weighting factor vectors obtained from the learning data of the Hidden-Markov model before learning, and the value is used at the time of pattern recognition.
The pattern recognition device described. 4. A pre-load coefficient vector calculation means is provided,
The weighting factor is stored in the weighting factor vector storage means as an average of the weighting factor vectors obtained from the learning data of the Hidden-Markov model before learning the Hidden-Markov model, and the value is used during the learning of the Hidden-Markov model. Item 2. The Hidden Markov model creation device according to item 2.