JPH06230795A - Voice recognizing method using basic hidden markov model - Google Patents

Abstract

PURPOSE:To provide the voice recognizing method using basic hidden Markov model(HMM) of high-speed, high recognition accuracy. CONSTITUTION:This voice recognizing method is provided with a learning processing part 6 for analyzing the distribution of continuous time in each state by learning the output probability and transient probability of the basic HMM from a voice analysis part and learning data, parameter storage part 7 storing the output probability, transient probability to be used as an initial value when starting the learning of the basic HMM, output probability, transient probability outputted from the processing part 6, distribution of continuous time in each state, minimum continuous time and maximum continuous time, recognition processing part 9 for performing continuous time control according to the transient probability different at every resident time by using the transient probability at every resident time and the output probability taken out of the storage part 7, and recognized result output part 10 for outputting the recognized result according to the generation probability of respective candidates, the voice recognizing method estimates the transient probability at every resident time so that the distribution of continuous time in each state calculated from the transient probability at every resident time can be coincident with the distribution of continous time in each state at the time of learning extracted from the storage part 7.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method using a basic hidden Markov model, and more particularly to a speech recognition method using a basic hidden Markov model for improving recognition accuracy by controlling duration.

[0002]

2. Description of the Related Art In the field of speech synthesis and recognition, when considering a voice signal, attention is paid to how small a certain voice lasts, and this last minute time is expressed as a duration. ing. By the way, in the conventional speech recognition method using a basic Hidden Markov Model (HMM), this duration is controlled for various reasons. The method of this duration control uses an HMM that has a structure that reflects the duration distribution of states, which is the so-called Markov process. At the time of learning the transition probability for each duration (residence time) for each state, Besides estimation, transition probabilities and output probabilities, there are other methods that introduce a penalty for giving a high score to a thing whose duration is a desired length in the calculation of the occurrence probability.

[0003]

In the above methods (1) and (2), the number of transition probabilities to be estimated at the time of learning is increased to increase the amount of calculation at the time of learning and the number of data per parameter. There is a problem that a large amount of learning data is required in order to prevent the estimation accuracy from decreasing due to the decrease of the estimation accuracy. That is, in the basic HMM, the transition probability is estimated for one state by the number of transitions leaving the state, but in these and these methods, the transition probability proportional to the stay time to be considered must be estimated. In the end, the number of transition probabilities to be estimated at the time of learning is several times that of the basic HMM without the duration control.

Then, a basic HMM without controlling the duration time.
In, the calculation at the time of speech recognition is to calculate the sum of the occurrence probability of the previous state multiplied by the output probability and the transition probability, but the method of You have to calculate the sum of things. Therefore, in the method of, in addition to the calculation of the transition probability and the output probability performed in the basic HMM without the duration control,
There is a problem that the amount of calculation at the time of recognition increases because it is necessary to calculate the penalty that gives a higher score as the stay time in the state subject to the calculation of the current transition probability and the output probability is the desired length. is there. Another problem is that there is no guarantee that the recognition rate will always improve because the method of incorporating the penalty into the occurrence probability is based on experience.

The present invention provides a speech recognition method using a basic hidden Markov model that solves the above-mentioned problems.

[0006]

SOLUTION: To solve the problem, an input voice signal is analyzed, and a voice analysis unit 6 for calculating a characteristic parameter thereof is provided, and output probabilities and transition probabilities of a basic hidden Markov model are learned from learning data. Further, the learning processing unit 6 for analyzing the distribution of the duration of each state based on the result is provided, and the output probability and transition probability used as initial values at the start of learning of the basic hidden Markov model, and the learning processing. The output probability and the transition probability output from the unit, the distribution of the duration of each state, the minimum duration, and the maximum duration are stored in the parameter storage unit 7. The transition probability for each stay time and the parameter storage are provided. A recognition processing unit 9 is provided for executing recognition by a basic hidden Markov model in which duration control is performed with transition probabilities that differ for each stay time using output probabilities obtained from the parts. In the speech recognition method using the basic hidden Markov model, which includes the recognition result output unit 10 that outputs the recognition result based on the occurrence probability of each candidate calculated by the recognition processing unit 9, the calculation is obtained from the transition probability for each stay time. Recognition method using a basic hidden Markov model that estimates transition probabilities for each stay time so that the distribution of the duration of each state is extracted from the parameter storage unit to match the distribution of the duration of each state during learning , Was constructed.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION The speech recognition method using the basic HMM of the present invention will be summarized. The transition probability estimated by the learning of the basic HMM can be considered as an average value of the transition probabilities for each stay time obtained by the conventional method described above. In addition, the minimum and maximum durations can be determined based on the analysis result of the duration distribution at a constant risk rate. Therefore, the self-transition probability up to the minimum duration is 1.0 and the self-transition probability after the maximum duration is 0.0, and the transition probability between these two is manipulated so that the distribution of the duration becomes equal to the distribution obtained during learning. If set to, the duration control can be easily performed.

An embodiment of a voice recognition method using the basic HMM of the present invention will be specifically described with reference to FIG. In FIG. 1, 1 is an AD conversion unit, 2 is a voice section detection unit, 3 is a control unit, 4 is a characteristic parameter calculation unit, 5 is a vector quantization unit, 6 is a learning processing unit, 7 is a parameter storage unit, and 8 is A transition probability estimation unit, 9 is a recognition processing unit, and 10 is a recognition result output unit.

First, common processing will be described. When voice is input to the AD conversion unit 1, the voice input is band-limited and converted into digital data. At the same time, the voice section detection unit 2 distinguishes the voice section from the noise section and detects the end of utterance. The control unit 3 is notified of the end of utterance. When the control unit 3 receives the notification of the end of utterance, the characteristic parameter calculation unit 4 calculates the digital Fourier transform, the LPC cepstrum and other characteristic parameters for the input voice AD-converted by the AD conversion unit 1. The vector quantizer 5 vector-quantizes the feature parameter sent from the feature parameter calculator 4 and outputs the code sequence.

Next, the learning process will be described. Before starting learning, the learning processing unit 6 retrieves the initial values of the output probability and the transition probability of each HMM from the parameter storage unit 7 and initializes the HMM. When the learning is started, the learning processing unit 6 learns the HMM corresponding to the output of the vector quantization unit 5 according to the instruction of the control unit 3. The control unit 3 monitors the progress of learning, and ends learning of the output probability and transition probability of the HMM when it is determined that the learning has converged. Next, the path on the trellis is calculated using the learning result, the duration of each state is determined, the distribution of duration is analyzed, and the sum of the frequencies is normalized to be 1.0. When the above processing is completed, the learning processing unit 6 stores the result of the output probability, the transition probability, and the distribution of the duration in the parameter storage unit 7.

The recognition process will be described below. Prior to starting the recognition, information on the transition probability of each HMM and the duration distribution for each state is taken out from the parameter storage unit 7 based on the command of the control unit 3, and these are calculated by the transition probability estimation unit 8
Send to. The transition probability estimating unit 8 stays so that the distribution of the duration of each state obtained by calculation from the transition probability of each stay time matches the distribution of the duration of each state at the time of learning extracted from the parameter storage unit 7. Adjust and estimate the transition probability for each time.

The distribution of the calculated duration used here is obtained as follows. Here, the self-transition probability is the probability of staying in the same state without making a transition from a certain state, so the sum of the probability of making a transition to its own state and the probability of making a transition to another state is 1.0. Therefore, the self-transition probability is 1.0
Subtracted from, this is the probability of leaving the current state. From this, the probability of transition to another state after performing self-transition n times is given by the following equation. That is, the frequency of the duration n (when reaching a certain state at time 0 and transiting to another state at time n + 1) is the product of the self-transition probabilities from the stay time 1 to the stay time n (1-stay It is the product of the self-transition probability of time n + 1). By the way, in general, when the distribution is obtained, when the number of stay times is how many, the expression “frequency” is adopted. In this case, it is not possible to count the actual number, but it is possible to know what is the probability of each stay time. Therefore, this probability is used instead of frequency to calculate the frequency distribution. Frequency of duration n = Self-transition probability of dwell time 1 x Self-transition probability of dwell time 2 ... …… Self-transition probability of dwell time n × (1-Self-transition probability of dwell time n + 1) Duration n is minimum continuation Calculate the frequency in the range from time to maximum duration. Normalize the sum of frequencies to 1.0. When the transition probability for each stay time is calculated, the control unit 3
Let us know. Then, the control unit 3 causes the parameter storage unit 7 to
The output probability is sent from the transition probability estimation unit 8 to the recognition processing unit 9 at each continuation time.

When a code sequence is sent from the vector quantization unit 5 to the recognition processing unit 9 based on a command from the control unit 3 at the time of speech recognition, the recognition processing unit 9 causes each H of the input code sequence.
Likelihoods for the MM are obtained, and these likelihoods are sent to the recognition result output unit 10 in the order of likelihood magnitude. The control unit 3 outputs to the recognition result output unit 10 a candidate having a certain likelihood or more.

In learning, according to the present invention, the transition probabilities estimated from the use of the basic HMM without the duration control may be one set for each state. In the method of using the HMM having a structure reflecting the state duration distribution, the number of states increases by the number of durations to be considered, and thus the number of transition probabilities to be estimated increases in proportion to the number of states. In the method of estimating the transition probability for each stay time in each state at the time of learning, it is necessary to estimate the transition probability by the number of durations to be considered. When the number of transition probabilities to be estimated increases, the number of learning data for estimating one transition probability decreases accordingly, so that the estimation accuracy decreases accordingly. In general, when estimating a certain number, the larger the number of data used, the higher the estimation accuracy can be. The higher the estimation accuracy, the higher the recognition rate. If the calculation amount is small, the same computer can naturally perform high-speed calculation.

In the case of speech recognition, the present invention is similar to that in the method of estimating the transition probability for each stay time in each state at the time of learning. When the amount of calculation is compared with the method of using an HMM having a structure that reflects the duration distribution of states, the amount of calculation is small as the number of states is small. In addition to the transition probabilities and output probabilities, when comparing the method that introduces a penalty that gives a high score to those whose duration is a desired length into the calculation of the occurrence probability, the minute calculation that does not need to calculate the penalty is It can be small.

[0016]

As described above, according to the present invention, the amount of calculation is small and the estimation accuracy of parameters is high as compared with the conventional methods 1 and 2, so that the voice recognition is performed at high speed and with high recognition accuracy. There is an effect that can be.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an embodiment of the present invention.

[Explanation of symbols]

 1 AD conversion unit 2 Voice section detection unit 3 Control unit 4 Characteristic parameter calculation unit 5 Vector quantization unit 6 Learning processing unit 7 Parameter storage unit 8 Transition probability estimation unit 9 Recognition processing unit 10 Recognition result output unit

Claims (1)

[Claims] 1. An input speech signal is analyzed, and a speech analysis unit for calculating a characteristic parameter of the speech signal is provided, learning of output probabilities and transition probabilities of a basic hidden Markov model is performed from learning data, and the result is further analyzed. A learning processing unit that analyzes the distribution of the duration of each state based on the output probabilities and transition probabilities used as initial values at the start of learning of the basic hidden Markov model, and the output output from the learning processing unit. Probability and transition probability, distribution of duration of each state,
A parameter storage unit that stores the minimum duration time and the maximum duration time is provided, and the duration control is performed by the transition probability for each stay time and the transition probability that is different for each stay time using the output probability obtained from the parameter storage unit. It uses a basic hidden Markov model that has a recognition processing unit that executes recognition by the basic hidden Markov model and that has a recognition result output unit that outputs the recognition result based on the occurrence probability of each candidate calculated by the recognition processing unit. In the speech recognition method, the distribution of the duration of each state calculated from the transition probabilities for each duration of stay is adjusted so that it matches the distribution of the duration of each state during learning extracted from the parameter storage unit. A speech recognition method using a basic hidden Markov model characterized by estimating transition probabilities.