KR20050014183A - Method for modificating state - Google Patents

Abstract

PURPOSE: A state probability correction method is provided to increase recognition performance of training and speech data consistently by training an HMM(Hidden Markov Modeling) state weight value by using an extended linear loss function in an HMM based speech recognition system. CONSTITUTION: A pattern of speech inputted through a discriminant function is recognized(S110). A class of the inputted speech is determined(S120). An average loss function that indicates speech recognition performance according to the determination of the class is obtained(S130). A model parameter equation is derived by the average loss function(S140). A state weight value training is performed according to the model parameter equation(S150). A weight value of an optimal state for time is applied to a probability density component in the discriminant function in each class.

Description

State Probability Correction Method {METHOD FOR MODIFICATING STATE}

The present invention relates to a state probability correction method, and more particularly, to improve speech recognition performance by training state weights using an extended misperception measure to prevent overadaptation of state weights to training data in an MCE training process. A state probability correction method using a determined loss function.

In general, Hidden Markov Modeling is widely used to express speech features in speech recognition. Hidden Markov modeling results in modeling the features of speech on a statistical basis with a probability distribution estimation method of the HMM state. The most widely used method for estimating the state probability distribution is ML (Maximun Likelihood) estimation method.

However, the ML estimation method is very difficult to find out accurate information about the probability distribution of the training data, and it is almost impossible to model the information distinguished between similar voices. Discriminant training methods have been studied as a way to make up for the shortcomings of the probability distribution estimation in the pattern recognition problem.

In particular, the performance of the speech recognizer is defined as the average recognition error rate of the recognizer and the optimum recognizer is the recognizer exhibiting the minimum recognition error rate. The classification training technique that has been mainly studied in this respect is the Minimum Classification Error (MCE) training method based on the Generalized Probabilistic Descent (GPD) algorithm.

The purpose of the MCE training technique is not to obtain a model by estimating the probability distribution of the data, but to distinguish the observed data of the HMM for the best recognition results. The following describes the MCE algorithm briefly and discusses an optimization category that represents a reasonable estimate of the error probability. In addition, the MCE training method ultimately approaches from the point of view of minimizing cognitive errors and adjusts HMM basic parameters such as feature extraction, acoustic modeling techniques, and mixture weights, averages, and standard deviations with fixed precision of the acoustic model. It has already been studied that performance improvement of speech recognition can be achieved by doing so. As an extended method of the MCE training method, a method of introducing state weights to optimize the recognizer using information distinguished between voices inherent in HMM state probability has been studied. The MCE training technique is mainly performed in parallel with the ML training technique and shows better performance than the HMM estimated by the ML training technique.

In a basic HMM-based speech recognizer, the classification function for class i for pattern recognition is defined as

From here, Is the optimal state string that maximizes the joint state-observed probability function for class i, Denotes a state transition probability of transitioning from state i to state j.

Is an observation vector in state j Represents a probability density function. In a continuous multivariate mixture Gaussian HMM, the state output distribution is

Where N (·) denotes the multivariate Gaussian density function Is the mean vector in state j, mixture m, Is the correlation matrix in state j, mixture m.

For the input voice, a class decision rule is used and the class C (X) for the input voice X is determined by the following rule.

From here, Denotes the class determined for the input speech or observation vector sequence by the classification function.

First, it is necessary to express the mathematical decision formula (3) in the form of a function. A class misrecognition measure that implies a decision rule as a continuous function of the parameter set Λ of the pattern recognizer is defined as follows.

Where η is a positive constant and N is the number of N-best misrecognition classes. For voice X corresponding to class i, Means misrecognition Means accurate recognition.

The complete loss function is defined in terms of misrecognition measures in the form of a smooth binary loss function.

The smooth binary loss function can be defined as any continuous binary function, but usually the following sigmoid function is used.

Where θ is set to zero or somewhat less than zero and r is a constant value.

Finally, recognizer performance for unknown speech is estimated as the average loss function.

Where l (·) is an indicator function.

The optimal model parameter is a model parameter that minimizes the average loss, and the GPD algorithm is mainly used to minimize the average loss. The GPD algorithm is given by

Where U is a matrix defined as positive, Is the step size of the learning rate or adjustment, Is a set of model parameters at time n.

GPD algorithm is an optimization technique without constraints. However, some constraints must be given to maintaining the HMM structure as a probabilistic model. Instead of using a complex constraint GPD algorithm, the GPD algorithm is applied to the transformed HMM parameters. The parameter conversion process does not have any constraints in the transformed space and the HMM constraint must be satisfied when converting to the raw space. Constraints on the following HMM parameters must be maintained in raw space:

In order to satisfy the parameter constraints in the raw space as in the above-described equation, a parameter conversion process as in the following equation is used before and after parameter training.

However, MCE training based on the segmental GPD algorithm requires the process of finding the optimal state sequence by the training class and the longitude calculation for the parameters constituting the HMM. There is a problem that it is difficult to consistently improve the speech recognition performance with respect to the training data and the recognition data.

The present invention devised to solve the above problems, the training that occurs in the process of introducing a weight corresponding to the state of each HMM constituting the set of speech recognition units to reduce the recognition error for the training data and training it State probabilities that can improve speech recognition performance for training data and recognition data by mitigating overadaptation problems for training data by training state weights by introducing extended loss functions to solve overadaptation problems with data. The purpose is to provide a correction method.

1 is a flowchart illustrating a state probability correction method according to an embodiment of the present invention;

2 is a graph showing state weight training using SIG,

3 is a graph showing state weight training using SIG + WL;

4 is a graph showing state weight training using LIN,

5 is a graph showing state weight training using LIN + WL.

In order to achieve the above object, the state probability correction method of the present invention includes: recognizing a pattern of a voice input through a classification function; Determining a class for the input voice; Obtaining an average loss function representing speech recognition performance according to the class determination; Deriving a model parameter equation by the average loss function; And performing state weight training according to the model parameter equation, wherein the classification function applies a weight of an optimal state with respect to time in each class to a probability density component in the classification function.

Hereinafter, the most preferred embodiments of the present invention will be described with reference to the accompanying drawings so that those skilled in the art can easily implement the technical idea of the present invention. .

1 is a flowchart illustrating a state probability correction method according to an exemplary embodiment of the present invention. The state probability correction method of the present invention will be described below.

First, the pattern of the voice input through the classification function is recognized (S110). In this case, in order to use the distinguishing information inherent in the HMM state output probability, each state has a weight and each state output score is expressed as a product of the state output probability value and the state weight. Suppose we have M HMMs as basic speech recognition units and each recognition unit consists of J states. The basic recognition unit is usually a phoneme or word model. In addition, in the conventional HMM, the classification function for class i in the state weighted HMM is defined by the following equation, similarly to the classification function for class i.

From here, Is the weight of the optimal state for time t in class i.

Thereafter, the class decision rule is applied to the input voice (S120). That is, the class is determined for the input speech or the observation vector sequence.

Thereafter, an average loss function representing speech recognition performance is obtained according to the misperception of the class (S130).

Thereafter, a model parameter equation is derived from the average loss function (S140). That is, the training of the state weight set to the initial value 1 is expressed by the model parameter equation as follows through the same process as the above Equation 4 to Equation 8.

Similar to equation (9), constraints on state weights must be established in raw space. That is, the state weight in the speech recognition unit should be constrained by the following equation.

From here, Is the state weight for state j in the HMM that is the basic recognition unit and J is the total number of states in the recognition unit.

Finally, state weight training is performed according to the model parameter equation (S150), and parameter conversion is performed before and after the state weight training, wherein the parameter transformation is given by the following equation.

From here, Is Is the state weight in the transformed parameter space.

In addition, the initial value of the training of the state weight may be set to one.

In general, MCE training uses parametric adaptive training using misrecognized voice, so if the amount of training data used for MCE is too small, an overadaptation problem with the training data may occur. The difference in recognition performance occurs. Therefore, a method for overcoming the difference in the recognition performance of the training data and the recognition data is required. For this purpose, an alternative loss function is considered instead of the smooth binary loss function. To do this, in the process of calculating the misperception scale, weighting the probability value of the string model of the training voice and adding it to the misperception scale can extend the misperception scale and select it as an extended linear loss function. We can choose the S-type (sigmoid) binary loss function by substituting An extension to the misperception scale that adds weights to the probability values of the strings of training speech is defined as follows.

Here, k is a coefficient of the probability value added further, and increases the recognition error of the training data and increases the derivative value of the HMM parameter of the probability value for the string model of the training voice.

In order to evaluate the performance improvement of speech recognition through state probability correction of HMM, experiments using single digits were conducted. The single digit DB was recorded in a quiet office environment and consisted of 500 voices (250 males, 250 females), 400 voices (200 males, 200 females) and 100 voices (male) 50 voices and 50 female voices were used for recognition. In this experiment, to sample the speech recognizer on the 11.025KHz DSP board, we sampled the 16KHz downsampled to 11.025KHz. As the feature vectors, a total of 3 streams and 39th feature vectors were constructed by using 13th order cepstral coefficients, 13th order first derivatives, and 13th order second derivatives. All state output probability distributions used an 8-mixture multivariable Gaussian distribution. As a baseline experiment, a model consisting of three states, three streams and eight mixtures was generated by the ML training method. In training HMM weights, the three most competitive strings were used to calculate the misperception scale (N = 3).

In this experiment, we use MCE training of HMM basic parameters using four loss functions (binary loss function due to misperceived scale, binary loss function due to extended misperceived scale, misrecognition scale, extended misunderstanding scale). Was carried out. Table 1 below shows word recognition rates for training and recognition data when the training data recognition rate is maximized. As the training is repeated, the training data recognition rate gradually increases, but the recognition rate for the recognition data is almost unchanged, which means that the state probability distribution for the ceptral coefficient is over-adapted to the training data.

In order to evaluate the performance of acoustic modeling with HMM state weights in HMM-based speech recognizer, we compared and evaluated the change of recognition performance by repeatedly training using the four loss functions mentioned above. In the training process, the state weight was adjusted through the parameter transformation process to maintain the HMM structure as a probabilistic model. Table 2 below shows the recognition rate for the training data and the recognition data when the word recognition rate for the recognition data is maximum.

Figure 2 shows the change in the recognition rate of the training and recognition data with repeated training using the S-type (sigmoid) loss function. As the training is repeated, the recognition rate of the training data tends to increase initially, but the recognition rate decreases due to the overadaptation of the training data used for MCE training. FIG. 3 shows a change in recognition rate according to training by obtaining an extended misperception scale by weighting the probability values of training phoneme strings for training speech in the case of k = 0.005, and obtaining a sigmoid loss function therefrom. As it is repeated, the recognition data recognition rate oscillates as well as the recognition rate of the training data decreases.

4 shows that the result of training the state weight using the linear loss function in the case of k = 0 does not show a consistent recognition performance improvement for the training data and the recognition data. FIG. 5 shows that the results of training using the extended linear loss function obtained by weighting the probability values of the trained phoneme strings for the training speech in the case of k = 0.005 show a consistent improvement in recognition performance not only for the training data but also for the recognition data. Indicates.

When the recognition rate of the training data and the recognition data is changed according to the weight multiplied by the probability value of the training phoneme sequence model in the linear loss function, the recognition results when the recognition rate for the recognition data is maximized are shown in Table 3 below.

Using the extended misperception scale in addition to the misperception scale by giving the weight k to the probability value of the training phoneme model for training speech has the effect of increasing the recognition error for the training data, and also the linear loss function for the state weight. It is effective to increase the derivative of by (1 + k) times. Therefore, by appropriately adjusting the k value, it is possible to obtain a result in which the training data and the recognition data recognition rate are improved consistently.

The present invention described above is capable of various substitutions, modifications, and changes without departing from the technical spirit of the present invention for those skilled in the art to which the present invention pertains. It is not limited to the drawings shown.

In the HMM-based speech recognition system, the HMM state weights are trained using the extended linear loss function obtained by weighting the probability values of the training phoneme sequences for the training speech so that the recognition performance of the training and speech data can be consistently increased. There is an advantage.

Claims (6)

Recognizing a pattern of speech input through the classification function; Determining a class for the input voice; Obtaining an average loss function representing speech recognition performance according to the class determination; Deriving a model parameter equation by the average loss function; And Performing state weight training according to the model parameter expression Including, The classification function is to apply the weight of the optimal state with respect to time in each class to the probability density component in the classification function. State probability correction method, characterized in that. The method of claim 1, Obtaining an average loss function representing the speech recognition performance according to the determination of the class, Using the extended misperception scale obtained by weighting the probability value of the string of training speech and adding the misperception scale State probability correction method, characterized in that. The method of claim 1, The initial value of the training of the state weight is set to 1 State probability correction method, characterized in that. The method of claim 3, The model parameter expression is given by the GPD algorithm. State probability correction method, characterized in that. The method of claim 4, wherein The state weights in the speech recognition unit, which are phonemes or words, have constraints established in raw space. State probability correction method, characterized in that. The method of claim 3, Parameter transformation by the model parameter expression is performed before and after the training of the state weight. State probability correction method, characterized in that.