JP3364234B2 - Pattern recognition device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an HMM (Hidden Markov) capable of improving a pattern recognition rate.
Model). [Summary of the Invention] [0002] The present invention relates to a pattern recognition apparatus for performing time-series pattern recognition using an HMM, and obtains a function representing the possibility of occurrence of a recognition error from all given learning data. By calculating an HMM parameter that minimizes the following, pattern recognition using an HMM having a higher recognition ability than in the past can be performed. [0003] Conventional HMM parameter learning methods include:
For example, the following three learning methods are known. (1) Baum-Welch algorithm This is also called the Forward-Backward algorithm, and is a widely used technique. In this method, when a state transition probability A and an output probability B, which are HMM parameters, are given, the maximum likelihood L (O | A, B) that the HMM outputs learning data O is maximized. This is a method of estimating a new state transition probability A and an output probability B based on the estimation. By repeating this parameter estimation,
It converges HMM parameters to local optimal values (eg, LRRabiner, BHJuang, "An Introduct
ion to Hidden Markov Models ", IEEE ASSP MAGAGINE, JA
N.1986, pp.4-16). (2) Error correction learning In this method, learning data is first recognized after HMM parameter estimation by the Baum-Welch algorithm,
When the likelihood difference is small even if misrecognized or correctly recognized,
Modify HMM parameters. That is, for the true category, the frequency of the symbol (label) of the training data is calculated as the output probability B
In addition to the above, a correction is made such that subtraction is performed from an erroneous category or a category having a small likelihood difference. Thereby, the recognition performance for learning data is improved (for example, LRBahl, PFBrown, PVde Souza, and RLMerce
r, "A New Algorithm for the Estimation of Hidden Ma
rkovModel Parameters ", Proceeding of the 1988 IEEE
International Conference on Acoustics, Speech and S
ignal Processing, pp493-496). (3) HMM learning by annealing In this method, an energy function having a negative logarithmic likelihood, which is created only from the learning data of the own category, is reduced by an annealing method. The HMM parameter is the state transition probability A
And the output probability B are selected alternately, and a random number according to a Gaussian distribution is added (for example, Douglas B. Paul, "Training of HMM Recogniz
ers by Simulated Annealing ", Proceedings of the 198
5 IEEE Internatinal Conference on Acoustics, Speec
h, and Signal Processing, pp13-16). [0007] However, (1)
Baum-Welch algorithm shown in (3)
In the learning of the HMM by annealing shown in (1), since learning data of other categories is not considered, these methods are not necessarily learning methods for increasing the degree of separation between categories. When the Baum-Welch algorithm of (1) is used, the HMM parameters converge to a local optimum value depending on the initial value, but the convergence value is not always the optimum value. On the other hand, the error correction learning shown in (2) considers learning data of other categories, and shows a very good result with respect to learning data, but recognizes too little with respect to unknown data. No improvement in rate is seen. This is because the error correction learning corrects the HMM parameters only between the true category and the incorrect category or the category having a small likelihood difference, so that it cannot be said that the correction is made in the direction in which the entire category can be easily recognized. Because. In addition, since the data is too adapted to the learning data, the recognition performance does not increase for unknown data when a category with many variations is handled. The present invention has been made in view of the above circumstances, and has as its object to obtain a function representing the possibility of occurrence of a recognition error from all learning data for all categories and calculate an HMM parameter for minimizing the function. By doing
An object of the present invention is to provide a pattern recognition apparatus using an HMM having a higher recognition ability than a conventional method. In order to achieve the above object, a pattern recognition apparatus according to the present invention comprises: a setting means for setting learning data and initial HMM parameters when performing pattern recognition using an HMM; and the initial objective function representing the likelihood that the initial recognition error from HMM parameter occurs, nearest category for each learning data
Using the sigmoid function
Function calculation means for obtaining the sum by category over the entire category, and perturbing the HMM parameters to obtain a new H
Process for obtaining MM parameters, new HMM obtained
In the process of obtaining a new objective function using the parameters, and when the obtained new objective function is decreasing,
HMM parameter calculation means for obtaining a HMM parameter that minimizes the objective function by appropriately repeating a process of adopting the new HMM parameter and the new objective function,
It is characterized in that pattern recognition is performed using the HMM parameters. (A) Set learning data and initial HMM parameters. (B) An initial objective function representing the possibility of a recognition error occurring is obtained from all the set learning data and initial HMM parameters. (C) Perturbation is applied to the HMM parameters to obtain new HMM parameters. (D) A new objective function is obtained using new HMM parameters. (E) If the objective function has decreased, a new HMM parameter and objective function are adopted. (F) The above steps (C), (D) and (E) are repeated as appropriate to determine HMM parameters for minimizing the objective function. According to the present invention having the above structure, the HMM parameter for minimizing the objective function representing the possibility of occurrence of a recognition error can be obtained, and the recognition rate when the present invention is applied to speech recognition is improved. . An embodiment of the present invention will be described below with reference to the drawings. First, the notation of this embodiment is defined as follows. M: the number of categories of the time series pattern to be recognized m, m ': category number (m, m' = 1, 2,...,
M) N: number of learning data in category n: learning data number (n = 1, 2,..., N) Omni: n-th learning data in category m Λ: set of HMM parameters of all categories, Λ = {λm} λm: HMM parameters category ^m, λm = {a m
^{ij, b m ijk} a m} ij: Category m, transition probability b ^m ijk of when the state of the HMM is changed from i to j: Category m, when the state of the HMM is changed from i to j, discrete output In the case of a distributed HMM, the probability L (Omn | λm) of outputting the symbol k: the likelihood Emn of outputting the learning data Omn from the category m: the difference in log likelihood between the nearest category and the subject category with respect to the learning data Omn E: Objective function that represents the possibility of recognition errors occurring in the entire category The present invention is characterized in that a set of HMM parameters Λ = {λm} that minimizes the objective function E is calculated. Here, a sigmoid function F (x) = 1 / (1 + exp (−x / μ)) (where μ is a constant) (3) is used as the function F, and the steepest descent method is used as a method for minimizing E. An embodiment using the method will be described with reference to the flowcharts shown in FIGS. In the flowcharts shown in FIGS. 1 and 2, steps ST1 and ST2 correspond to the setting means of claim 1, step ST3 corresponds to the objective function calculating means of claim 1, and steps ST4 to ST11 correspond to the HMM parameter calculating means of claim 1. Configure each. First, learning data is set (step S).
T1). In this method, time series pattern learning data is prepared for all categories. Next, a set 初期 of initial HMM parameters is set (step ST2). An arbitrary initial value can be set for the state transition probability and the output probability, which are the HMM parameters. In the present embodiment, a method of setting HMM parameters after executing the Baum-Welch algorithm several times is used, and this set of initial HMM parameters is denoted by Λ. Note that a method may be used in which the time-series learning data is equally divided by the number of states of the HMM, and an output probability is statistically obtained for each symbol. Next, an initial objective function E is obtained (step ST3). This is because the difference Emn of the log likelihood between the nearest category and the own category with respect to the learning data Omn is expressed by the above (1).
Obtained by the formula The obtained log-likelihood difference Emn is converted by a monotonically increasing function F, and the sum of all learning data is obtained based on the above equation (2). This is the initial objective function E
It becomes. Log likelihood is a forward algorithm or Vite
Although determined by the rbi algorithm, in this embodiment, a forward algorithm is used. Next, HMM parameters to be perturbed are selected (step ST4). In this method, a uniform random number is generated to determine which HMM parameter is to be perturbed from all parameters of all categories, and only one is determined. Next, a perturbation is given to the selected HMM parameter (step ST5). This processing is performed in step S using uniform random numbers.
The value of the HMM parameter selected at T4 is changed. For example, if the state transition probability a ^m ij is selected, [1, -
1] multiply the above uniform random number r by a constant δ.
In the present embodiment, the constant δ = 0.01. [0028] Are similarly when the output probability b ^m ijk is selected, perturbing as follows. [0029] Next, it is determined whether or not to use the perturbation given as described above (step ST6). Since the state transition probability and the output probability both take the value of a probability, if the value after perturbation is smaller than 0 or larger than 1, the process returns to step ST4 to set another HMM parameter for perturbation. select. Next, a new set of HMM parameters Λ 'is determined (step ST7). The state transition probability of the perturbed category m,
Alternatively, the output probability is adjusted to a value that satisfies the following probability condition. [0034] Is a new set of HMM parameters Λ ′. Next, a new objective function E 'is obtained (step ST8). Using the newly obtained set of HMM parameters Λ ′, a new objective function E ′ is obtained in the same manner as in step ST3. Next, a decrease in the objective function is determined (step ST9). If E ′ ≦ E, the result of the perturbation is adopted to
= E ', and the set of HMM parameters Λ = Λ' (step ST10). Otherwise, if E '> E,
E and し な い are not updated. When the perturbation is performed for a preset number of loops and when the objective function E becomes very small, for example, when the objective function E becomes smaller than 10 ^-8 , the perturbation is terminated (step ST11). The set of HMM parameters ＭＭ which minimizes the objective function thus obtained is adopted as the optimum set of HMM parameters. In the above embodiment, the discrete output distribution H
Although the example in which the MM is used has been described, in the continuous distribution HMM, the present invention can be applied to a case where perturbation is added to the average and variance of the output probability density by using random numbers. In the above embodiment, the steepest descent method is used to minimize the objective function E.
Other optimization techniques can be used. Although the sigmoid function is used here also for the monotone increasing function F (x), any function that satisfies the condition of F (x) may be used.
For example, the following function can be considered. ## EQU1 ## Next, an experimental example of a speech recognition process to which the method of the present invention is applied will be described with reference to the block diagram of FIG. Here, the present invention is applied to the problem of recognizing Japanese voiced plosive consonants / b /, / d /, / g / among discrete speech recognition problems using a discrete distribution type HMM. In this speech recognition processing, voiced plosive consonants / b /, / d /, / g / are each modeled by one HMM, and the parameters of each HMM are learned by the method of the present invention. Here, a discrete distribution type HMM without four skips and three states is used. The voice data for learning and recognition is
An important word uttered by a single speaker in the ATR database in units of phrases is used. In this experiment, the case where speakers were MAU (adult male) and FSU (adult female) was examined. First, the beginning and the end of each consonant data of the input voiced consonant are sampled at a sampling frequency of 15 kHz (block B20) with reference to the label assigned to the ATR database, and cut out. An LPC cepstrum analysis (block 21) was performed. The code book is composed of 17 consonants (/ p /, / t /, / k /, / ts /, / s /, / h /, / z /, / ch /, phonologically balanced words of each speaker). / s
h /, / b /, / d /, / g /, / r /, / w /, / y /, / m /, / n /) with a size of 256 (block B22) and an LPC cepstrum The result of the analysis is vector-quantized (block B2
3). Next, 300 data of each consonant are divided into 1
Data sets 1, 2, and 3 were created by dividing the data set into 00 pieces (block B24). An experiment in which HMM learning is performed on one data set and recognition is performed on the other two data sets is performed, and the average recognition error rate in each data set is used as a result. The initial HMM parameters at the time of parameter learning were set by executing the Baum-Welch algorithm until the log likelihood of the own category converged (about 45 repetitions). Next, HMM parameters were learned by the learning method of the present invention using the learning data set (block B25). The log likelihood of the recognition data set was calculated using the created HMM parameters (block B26) (block B27). By judging the log likelihood (block B28), the category which became the maximum among all the categories was determined as the recognition result. FIG. 4 shows the recognition result of the Japanese voiced plosive consonant by the speaker MAU, and FIG. 5 shows the recognition result of the Japanese voiced plosive consonant by the speaker FSU. Here, for comparison, the initial HMM parameters of the learning method of the present invention, ie, Bau
The recognition error rate after m-Welch learning and the recognition error rate when the same experiment is performed by error correction learning are also shown. In addition,
In each figure, open means when unknown data is recognized,
Close means when the learned data itself is recognized. As can be understood from FIG. 4, the recognition error rate with the initial HMM parameters for the unknown data of the speaker MAU was 12.9%.
0.1%. This is superior to the recognition error rate of 12.3% by the error correction learning method. Also, as understood from FIG. 5, regarding the unknown data of the speaker FSU,
The recognition error rate with the initial HMM parameters is 11.6%, and the recognition error rate by the error correction learning method is 11.5%, whereas the recognition error rate is 10.4% in the present learning method. Was remarkably reduced. From the above experimental results, it was confirmed that the HMM parameter learning method of the present invention can provide HMM parameters having higher recognition performance than the conventional learning method. As described above, according to the present invention, it is possible to provide a pattern recognition device having a higher pattern recognition rate than the conventional one.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart showing functions of the device of the present invention. FIG. 2 is a flowchart showing functions of the device of the present invention. FIG. 3 is a block diagram illustrating a speech recognition process to which the present invention has been applied. FIG. 4 is an explanatory diagram showing a speech recognition result of FIG. 3; FIG. 5 is an explanatory diagram showing a speech recognition result of FIG. 3;

Continuation of the front page (56) References JP-A-4-205389 (JP, A) JP-A-3-176781 (JP, A) Tsutomu Matsunaga, Ichiro Abe, Hiromi Kida, Character recognition using the simulated annealing method Dictionary optimization, IEICE Technical Report [Pattern Recognition and Understanding], Japan, July 12, 1990, PRU90-39, p. 79-84 Akio Ando, Kazuhiko Ozeki, Standard Pattern Learning Algorithm for Minimizing False Recognition Functions, Proceedings of the Acoustical Society of Japan, Spring Meeting, 1991, Japan, March 27, 1991, p. 205-206 Takashi Imai, Akio Ando, HMM Learning Method to Minimize Error Function Based on Log-Likelihood Difference, Proceedings of the Acoustical Society of Japan 1991 Autumn Meeting, Japan, October 2, 1991, p. 79-80 Takashi Imai, Akio Ando, HMM Learning Method to Minimize Error Function Based on Log-Likelihood Difference, IEICE Technical Report [Speech], Japan, December 19, 1991, SP91-87 , P. 49-56 Akio Ando, Kazuhiko Ozeki, Standard Pattern Learning Algorithm for Minimizing False Recognition Functions, IEICE Transactions A, Japan, April 25, 1993, Vol. J76-A, p. 580-588 Shinobu Mizuta, Kunio Nakajima, Optimal Discrimination Learning Method for Mixed Continuous Distribution HMM, Proceedings of the Acoustical Society of Japan Spring Meeting, 1990, Japan, March 28, 1990, p. 23 −24 (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/14 G10L 15/06

Claims (1)

(57) [Claims] [Claim 1] When performing pattern recognition by HMM,
Setting means for setting the HMM parameters of learning data and the initial, the initial objective function representing the likelihood that recognition errors occur from all learning data and initial HMM parameters are set, the self and nearest Categories for each training data
Using the sigmoid function
Means for calculating an objective function by obtaining a sum in a field; processing for obtaining a new HMM parameter by perturbing the HMM parameter; processing for obtaining a new objective function using the obtained new HMM parameter; If the new objective function has decreased, a new HMM
HMM parameter calculation means for obtaining an HMM parameter for minimizing the objective function by appropriately repeating the process of adopting the parameter and the new objective function, and performing pattern recognition using the HMM parameter. .