JPH096387A - Voice recognition device - Google Patents

Abstract

PURPOSE: To provide a voice recognition device excellent in discriminating ability between similar words from each other. CONSTITUTION: This device is provided with a feature extraction part 3 extracting the feature data from a word voice segmented from an inputted voice, a state number estimation part 4 estimating a state number for the word voice when the feature data are modeled from the feature data by a Markov model, a similar word judging part 5 judging whether or not the word similar to the word voice to be registered newly has been registered, a state number addition part 6 increasing the estimated state number, a learning part 7 applying the feature data to a word model and obtaining a Markov model parameter, a voice dictionary file 8 consisting of the Markov model parameter and a collation decision part 9 calculating likelihood for respective word models and deciding a recognition candidate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing word speech and outputting the recognition result.

[0002]

2. Description of the Related Art Prior to explaining a conventional voice recognition device, Hidd used in the conventional voice recognition device is used.
The principle of speech recognition by the en Markov Model (herein referred to as "Markov model") will be described.

The Markov model has N states S1 and S.
2, ..., SN, transitions the states one after another with a certain probability (transition probability) at regular intervals, and at that time, labels (feature data) one by one with a certain probability (output probability). It is to output.

Then, the speech is regarded as a time series of labels (feature data), and at the time of learning, a Markov model is created by uttering each word several times to model them, and at the time of recognition, the label of the input speech is recognized. Recognition is performed by searching for a Markov model that maximizes the probability (likelihood) of outputting a sequence.

A detailed description will be given below with reference to the drawings.
FIG. 5: is explanatory drawing of the Markov model in the conventional speech recognition apparatus. What is shown is the Journal of the Acoustical Society of Japan 4
This is a simple example of a Markov model shown in Volume 2, No. 12 (1986), "Speech Recognition Based on Hidden Markov Model". This Markov model is composed of three states and has two kinds of labels a and b. Output a label sequence consisting of only.

The initial state is S1, and the transition from S1 to S1 itself has a probability of 0.3 (at that time, the label a is output. Since the output probability of the label b is 0.0, it is not output). , And transition to S2 with a probability of 0.7 (at that time, the label a is output with a probability of 0.5 and the label b is output with a probability of 0.5).

From the state S2, the state transits to S2 itself with a probability of 0.2 (at that time, the label a or the label b is output with a probability of 0.3 or 0.7, respectively) or 0.8. Transition to the final state S3 with probability (at that time, the label b is output. Since the output probability of the label a is 0.0, it is not output)
It represents that.

Here, considering the probability (likelihood) that this Markov model outputs a label sequence (sequence of feature data) abb, the state sequence permitted by this Markov model is S
There are only two, 1S1S2S3 and S1S2S2S3,
Probability is 0.3 * 1.0 * 0.7 * 0.5 * 0.8 * 1.0 = 0.0840 0.7 * 0.5 * 0.2 * 0.7 * 0.8 * 1.0 = 0.0392. Both possibilities are possible, so 0.0840 total
It can be seen that this Markov model outputs abb with a probability of + 0.0392 = 0.1232.

By learning the Markov model for each word in advance and obtaining the transition probability of the state most suitable for each word and the output probability of the label at each state transition, the label sequence of an unknown word becomes When input, probability (likelihood) calculation is performed on each Markov model, and it is possible to know which word the Markov model is likely to output this label sequence, and thus recognition can be performed. The above is the principle of speech recognition by the Markov model.

FIG. 6 is an exemplary diagram showing the correspondence between the time series of the voice waveform and the feature data and each state of the Markov model in the conventional voice recognition device, and shows the correspondence relationship when "beginning" is uttered. There is. In this way, the Markov model is represented in the time series of the voice feature data with a small number of states of the phoneme of the word.

By the way, in a conventional speech recognition apparatus for recognizing a word speech using a Markov model, at the time of learning, for each word registered in the speech recognition apparatus, the number of states, which is as small as the number of phonemes of the word, is used as the phoneme spectrum. Seeking from changes, etc.,
The output probabilities of the feature data at each state transition and the transition probabilities between states are estimated by learning and modeled in a Markov model, and the input speech is applied to all these models at the time of recognition, and likelihood calculation is performed for recognition. Was there.

[0012]

By the way, when a word is registered at the time of learning using a conventional voice recognition device, for example, after the word "Sato" is registered, a word acoustically similar to that word is registered. If "Katou" or the like is to be registered, the number of states and the Markov model parameters are the same, so if it is registered as it is, it becomes difficult to identify both words at the time of recognition. Therefore, if a similar word has already been registered, ask the user to say something like "Katokacho" instead of "Katou", and then
I had to register. As described above, the conventional voice recognition device has a problem that it is often necessary to reword the user because the ability to identify similar voices is low.

Therefore, an object of the present invention is to provide a voice recognition device having an excellent ability to distinguish between similar words.

[0014]

A voice recognition device of the present invention comprises a voice input means for inputting a voice containing a word voice, a word voice cutout portion for cutting out only the word voice portion from the input voice, A feature extraction unit that extracts feature data from the cut out word voice, a state number estimation unit that estimates the number of states for the word voice when modeling from the feature data by a Markov model, and a word voice that is newly registered Similar word determination unit that determines whether the word that has been registered is already registered, a state number addition unit that increases the estimated number of states, a learning unit that applies feature data to a word model to obtain Markov model parameters, and a learned Markov model A voice dictionary file consisting of parameters, a matching determination unit that calculates the likelihood for each word model and determines a recognition candidate, and a recognition result And a determination result output unit for outputting.

[0015]

With the above configuration, by further increasing the number of states estimated by the state number addition unit, the features of the word are expressed finely, and the voice recognition device can easily identify the similar word. You can register,
As a result, it is possible to keep the frequency of rewording low so that the user can distinguish similar words.

[0016]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a functional block diagram of a voice recognition apparatus according to an embodiment of the present invention. In FIG. 1, 1 is a voice input means for inputting voice including word voice, and 2 is word voice. A word-speech cutout unit that cuts out only the word-speech part from the sound, 3 is a feature extraction unit that extracts feature data from the cut-out word voice, and 4 is the number of states for the word voice when modeling from the feature data by a Markov model State number estimating unit 5, a similar word determining unit for determining whether or not a word similar to the word voice to be newly registered is already registered, 6 is a state number adding unit for increasing the estimated state number, and 7 is a feature A learning unit that applies the data to a word model to obtain Markov model parameters, 8 is a voice dictionary file including the learned Markov model parameters,
Reference numeral 9 is a matching determination unit that calculates a likelihood for each word model and determines a recognition candidate, and 10 is a determination result output unit that outputs a recognition result.

FIG. 2 is a circuit block diagram of a voice recognition apparatus in one embodiment of the present invention. In FIG. 2, 11 is a microphone, 12 is a ROM (read only memory) for storing programs, and 13 is a program of the ROM 12. CPU (Central Processing Unit) that executes the
Reference numeral 13 is a RAM (writable memory) for temporarily storing information necessary for executing the program, 15 is a monitor for displaying processing status to the user, and 16 is a file device for storing information.

It should be noted that the voice input means 1 in FIG. 1 uses a microphone 11 to include a word voice cutout unit 2, a feature extraction unit 3, a state number estimation unit 4, a similar word determination unit 5, a state number addition unit 6, and a learning unit 7. The collation determination unit 9 is realized by the CPU 13 executing a program stored in the ROM 12 while exchanging data with the microphone 11, the ROM 12, the RAM 14, and the file device 16. The voice dictionary file 8 is stored in the file device 16, and the determination result output unit 10 is realized by the monitor 15.

FIG. 3 is a flowchart for registration in one embodiment of the present invention, and FIG. 4 is a flowchart for recognition in one embodiment of the present invention.

The operation when the word voice "Katou" is registered in the voice recognition apparatus according to the present embodiment having the above-described configuration will be described with reference to the flowchart of FIG. Note that, here, it is assumed that the acoustically similar word "Sato" to "Katou" is acoustically registered in the voice dictionary file 8.

First, in step 1, a voiced voice containing the word voice "Katou" is input from the voice input means 1. In step 2, the word voice cutout unit 2 cuts out the word voice "Katou" from the voiced voice including the word voice "Katou". This can be realized by the word voice cutout unit 2 detecting and removing a silent or low noise part before and after the word voice "KATO" by the power of the voice or the like.

In step 3, feature extraction is performed by a method such as obtaining an LPC cepstrum coefficient for the word voice "Katou" by a linear predictive analysis (LPC analysis) in the feature extraction unit 3. In step 4, the number-of-states estimation unit 4 estimates the number of states for the word voice from the feature data extracted from the word voice “Kato” in step 3. The number of states can be estimated by the Proceedings of the Acoustical Society of Japan (199
0.3) "Regarding the number of HMM states and the number of mixtures in continuous digit speech recognition".

In step 5, the similar word determination unit 5 determines whether or not a similar word of the word voice "Katou" already exists in the voice dictionary file 18. This determination is made by using a general DP matching technique, or by using the word voice "Katou" to recognize the voice recognition device,
It is realized by judging whether there is a recognition candidate or not. As a result of the determination, if there is a similar word, the process proceeds to step 6, and if there is no similar word, the process proceeds to step 7. here,
Since there is a similar word "Sato", the process proceeds to step 6.

At step 6, the state number adding section 6
Increase the number of states estimated in step 4 by a certain percentage. The percentage value is determined in advance by evaluating the voice recognition device while gradually changing the percentage value until similar words can be identified, for example, by increasing the estimated number of states by 10%. be able to. As a result, the number of states can be increased considerably compared to the number of states of the phoneme of a word.

In step 7, the learning unit 7 sets the characteristic data of the word voice "Katou" in step 5 or step 6.
Learned using a Markov model having the number of states obtained in (the number of states obtained in step 6 because there are similar words in this example), and the Markov model parameters of the transition probabilities between the states and the output probabilities of the feature data at the transition And the obtained Markov model parameters are stored in the voice dictionary file 8.

Further, in step 7, the feature data of the word voice "Katou" is learned to obtain the Markov model parameters, and the obtained Markov model parameters and the feature data are stored in the voice dictionary file 8. If there is a similar word, the feature data of the word voice “Sato” that has already been registered in the voice dictionary file 8 and is judged to be a similar word is read, and learning is performed again using the Markov model having the number of states obtained in step 6. Then, the obtained Markov model parameter is replaced with the already stored Markov model parameter of “Sato”.

Next, the operation for recognizing the word voice "Katou" will be described with reference to the flowchart of FIG. First, in step 11, a voiced voice including the word voice "Katou" is input from the voice input means 1. In step 12,
The word voice cutout unit 2 cuts out the word voice "Katou" from the voiced voice including the word voice "Katou". In step 13, the feature extraction unit 3 performs feature extraction for the word voice "Katou".

In step 14, the matching determination unit 9 uses the feature data of the word voice "Katou" to perform likelihood calculation on the Markov model parameters of each word model read from the voice dictionary file 8, and the word model with high likelihood is calculated. Is determined as a recognition candidate. In step 15, the determination result output unit 10 displays the recognition result to the user.

[0030]

According to the present invention, when it is determined that the words are similar to each other, the number of states in the Markov model is artificially increased to express the features in detail to enable the discrimination between the similar words. Therefore, the frequency of rewording the user is reduced, and the recognition accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a circuit block diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 3 is a flowchart at the time of registration in one embodiment of the present invention.

FIG. 4 is a flowchart at the time of recognition in one embodiment of the present invention.

FIG. 5 is an explanatory diagram of a Markov model in a conventional speech recognition device.

FIG. 6 is an exemplary diagram showing correspondence between a voice waveform, a time series of characteristic data, and states of a Markov model in a conventional voice recognition device.

[Explanation of symbols]

 1 voice input means 2 word voice cutout unit 3 feature extraction unit 4 state number estimation unit 5 similar word determination unit 6 state number addition unit 7 learning unit 8 voice dictionary file 9 collation determination unit 10 determination result output unit

Claims (2)

[Claims] 1. A voice input means for inputting a voice including a word voice, a word voice cutout unit for cutting out only a portion of the word voice from the input voice, and a feature extraction for extracting feature data from the cut out word voice. Section, a state number estimation section that estimates the number of states for the word speech when modeling with the Markov model from the feature data, and similarity that determines whether a word similar to the word speech to be newly registered has already been registered. A word determination unit, a number-of-states addition unit that increases the number of estimated states, a learning unit that applies feature data to a word model to obtain Markov model parameters, and a voice dictionary file that includes the learned Markov model parameters.
A speech recognition apparatus comprising: a matching determination unit that calculates a likelihood for each word model and determines a recognition candidate; and a determination result output unit that outputs a recognition result. 2. In order to perform learning again for the registered words with the increased number of states, in addition to the learned Markov model parameters, the characteristic data of the word sounds are also stored. The voice recognition device according to claim 1, wherein the voice recognition device is a voice recognition device.