JPH0997095A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the speech recognition device which uses HMM(Hidden Markov Model) that enables a rejection threshold for handling the input of an incorrect spoken sound such as a given cough to be set and used with likelihood corresponding to a user. SOLUTION: In addition to a speech input means 1, a word voice segmentation part 2, a feature extraction part 3, a state estimation part 4, a learning part 5, etc., a likelihood output part 6 which finds likelihood from feature data and HMM parameters and a threshold setting part 8 which sets the threshold for rejection are constituted. Therefore, such trouble that a voice is rejected no matter how many times the user voices a word is eliminated and the usability of the user can be improved.

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 The conventional Hidden Markov
In order to describe a voice recognition device that recognizes a word voice using Model (abbreviated as HMM in the present invention), a method of voice recognition by HMM will be described first. The HMM has N states S1, S2 ,. . . , SN, and the state transitions one after another with a certain probability (transition probability) at regular intervals, and at that time, outputs a label (feature data) one by one with a certain probability (output probability). Is.

When the speech is regarded as a time series of labels (feature data), each word is uttered several times during learning to create an HMM that models them, and at the time of recognition, the label series of the input speech is recognized. H that maximizes the probability (likelihood) of outputting
Recognition is performed by searching for MM. Hereinafter, a specific description will be given with reference to the drawings.

FIG. 5 is a diagram showing an example of a conventional HMM, which is shown in "Hidden", Vol.
3 is a simple example of an HMM shown in "Voice Recognition Based on Markov Model". This HMM is composed of three states and outputs a label series consisting of only two types of labels a and b. The initial state is S1, and from S1,
Transition to S1 itself with a probability of 0.3 (at that time, label a
Is output. The label b is not output because the output probability is 0.0) or transits to S2 with a probability of 0.7 (in that case, 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, whether the label a or b is 0.
Output with a probability of 3, 0.7) or transition to the final state S3 with a probability of 0.8 (in this case, output the label b. Since the output probability of the label a is 0.0, it is not output). It represents.

Here, considering the probability (likelihood) that this HMM outputs a label series (characteristic data string) abb,
The state sequences allowed in this HMM are S1S1S2S3 and S
There are only 1S2S2S3, and the probability of each is
0.3 * 1.0 * 0.7 * 0.5 * 0.8 * 1.0 =
0.0840 and 0.7 * 0.5 * 0.2 * 0.7 *
0.8 * 1.0 = 0.0392. Both possibilities are possible, so total 0.0840 + 0.0392 = 0.12
It can be seen that with a probability of 32, this HMM outputs abb.

Therefore, if the HMM for each word is learned in advance and the transition probability of the state most suitable for each word and the output probability of the label at each state transition are obtained, the label sequence of an unknown word is input. In such a case, if the probability (likelihood) is calculated for each HMM, it is possible to know which word the HMM is likely to output this label sequence, and it is possible to recognize it. The above is the method of voice recognition by the HMM.

Further, FIG. 6 is an example diagram showing the correspondence between the time series of the voice waveform and the characteristic data and the respective states of the HMM in the conventional voice recognition, and shows the correspondence when uttering "beginning". In this way, the HMM is represented in a state in which the number of phonemes of the word is small with respect to the time series of the voice feature data.

In a conventional speech recognition apparatus for recognizing a word speech using an HMM, at the time of learning, for each word registered in the speech recognition apparatus, the number of states having a small number of phonemes of the word is determined from a change in the phoneme spectrum or the like. Obtained, the output probability of the feature data in each state transition and the transition probability between states are estimated by learning and modeled in HMM, and at the time of recognition, the input speech is applied to all these models to perform likelihood calculation,
I was aware.

[0009]

In the voice recognition device, in order to deal with the case where an unfair vocalization input such as throat clearing is received, the candidate with the highest likelihood is not always returned to the user, If the candidate with the highest likelihood does not exceed a certain threshold, it will be rejected and it will be important to prompt the user to speak again in order to improve operability. . However, the threshold for this rejection is a constant value determined by the voice recognition device provider in advance by evaluating the voice recognition device. Sometimes candidates were rejected and could not be recognized. By the way, according to Sadateru Furui's "Digital Signal Processing" (Tokai University Press), Chapter 10, 10.2 Speech recognition problem, "a small percentage of the total speakers, but extremely high recognition rate There is a problem with low speakers. "

This is because, for a speaker whose recognition rate is low, the feature data is out of the range of the average feature data of the speaker due to factors such as a difference in the clarity of voice peculiar to an individual and a fluffiness. This is because the likelihood is calculated lower than usual (in the sense that it is unlikely to occur probabilistically), and therefore the likelihood threshold set based on the average speaker feature data is set. There was always a phenomenon of not exceeding.

Therefore, according to the present invention, an HMM for recognizing a word voice that allows setting and using a reject threshold for responding to an input of an uttered voice such as a cough clearing with a likelihood according to a user. An object of the present invention is to provide a voice recognition device using the.

[0012]

To this end, the voice recognition apparatus of the present invention comprises a voice input means for inputting a voice including a word voice, and a word voice for cutting out only the word voice portion from the voice including the word voice. A cutout unit, 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 HMM, and applies the feature data to the word model. A HMM parameter learning unit, a likelihood output unit that calculates a likelihood from the feature data and the HMM parameter, a voice dictionary file including the learned HMM parameter and likelihood information, and a threshold value for rejecting are set. A threshold setting unit, a matching determination unit that calculates the likelihood for each word model and determines a recognition candidate, and a determination that outputs a recognition result And a result output unit.

[0013]

When the word is registered in the voice recognition device, the HMM parameter obtained by learning is used to recognize the voice input for registration, and the likelihood at that time is obtained. That is, even in the case of a speaker whose recognition rate becomes low, the (lower) likelihood corresponding to the utterance of the speaker is obtained as the likelihood of the user. Then, the likelihood is also registered in the voice dictionary file together with the HMM parameter. At the time of recognition,
The likelihood information in the voice dictionary file is read and used as a reference value for the likelihood threshold. Thereby, the likelihood threshold according to the user can be set and accurate recognition can be performed. Since it is possible to set the likelihood threshold according to the user in this way,
The problem that "the recognition candidate is rejected and cannot be recognized no matter how many times the user utters it again" does not occur.

[0014]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a configuration block diagram of a voice recognition device according to an embodiment of the present invention. In the figure, 1 is a voice input means for inputting a voice including a word voice, 2 is a word voice cutout portion for cutting out only a portion of the word voice from the voice including the word voice, and 3 is characteristic data extracted from the cut out word voice. A feature extraction unit, 4 is a state number estimation unit that estimates the number of states for a word voice when modeling with HMM from the feature data, 5 is a learning unit that applies the feature data to the word model to obtain HMM parameters, and 6 is a feature Likelihood output unit that obtains a likelihood from data and HMM parameters, 7 is a voice dictionary file that includes learned HMM parameters and likelihood information, 8 is a threshold setting unit that sets a threshold for reject, and 9 is The collation judging unit 10 which calculates the likelihood for each word model and judges the recognition candidate is a judgment result output unit which outputs the recognition result.

FIG. 2 is a circuit block diagram of a voice recognition device according to an embodiment of the present invention.
2 is a read only memory (ROM), 13 is a central processing unit (CPU), 14 is a writable memory (RAM),
Reference numeral 15 is a monitor, and 16 is a file device.

The voice input means 1 shown in FIG. 1 is collated with the word voice cutout unit 2, the feature extraction unit 3, the state number estimation unit 4, the learning unit 5, the likelihood output unit 6, the threshold value setting unit 8 by the microphone 11. In the determination unit 9, the CPU 13 has a microphone 11 and a ROM 12
By executing the program stored in the ROM 12 while exchanging data with the RAM 14 and the file device 16, the voice dictionary file 7 is realized by the file device 16 and the determination result output unit 10 is realized by the monitor 15.

FIG. 3 is a flow chart at the time of registration of the voice recognition device according to the embodiment of the present invention, and FIG. 4 is a flow chart at the time of recognition of the voice recognition device according to the embodiment of the present invention. A case where a certain word voice is registered in the voice recognition device configured as described above will be described based on the flowchart of FIG.

In step 1, the voice input means 1
A voiced voice including a word voice is input. In step 2, the word voice cutout unit 2 cuts out a word voice from a voiced voice including the word voice. This can be realized by detecting and removing silent or low noise portions before and after the word voice by the power of the voice or the like. In step 3, the feature extraction unit 3 performs feature extraction by a method such as obtaining an LPC cepstrum coefficient for the word voice by linear prediction analysis (LPC analysis). 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 in step 3. The number of states can be estimated by the Acoustical Society of Japan Proceedings (1990.
3) It can be performed based on "about the number of states and the number of mixtures of HMMs in continuous digit speech recognition".

In step 5, the HMM having the number of states obtained in step 4 for the feature data of the word voice by the learning unit 5
Learning is performed using the model, HMM parameters of transition probabilities between states and output probabilities of feature data at transitions are obtained, and the obtained HMM parameters are stored in the voice dictionary file 7. In step 6, the likelihood output unit 6 uses the feature data of the word voice to perform likelihood calculation on the HMM parameters read from the voice dictionary file 7 and found in step 5 to obtain the likelihood. Then, the likelihood information is also stored in the voice dictionary file 7.

Next, in the case of recognizing a certain word voice, this operation will be described below with reference to the flowchart of FIG. In step 11, the voice input means 1 inputs a voiced voice including a word voice. In step 12, the word voice cutout unit 2 cuts out the word voice from the voiced voice including the word voice. In step 13, the feature extraction unit 3 extracts the feature of the word voice. In step 14, the matching determination unit 9 performs likelihood calculation on the HMM parameters of each word model read from the voice dictionary file 7 using the feature data of the word voice, and determines a word model with a high likelihood as a recognition candidate.

In step 15, the threshold value setting unit 8 sets a threshold value for rejecting based on the likelihood information read from the voice dictionary file 7. This threshold value is set as "the read likelihood information is used as it is" or, "the threshold value determined by evaluating the voice recognition device is weighted with the read likelihood information. What has been done is used as a threshold ".
In step 16, the collation determining unit 9 determines in step 14
It is determined whether or not the likelihood of the recognition candidate obtained in step 15 exceeds the threshold value set in step 15. If it exceeds, the process proceeds to step 17. If it does not exceed, it is rejected and the user inputs it again. Return to step 11 to get the request.
In step 17, the determination result output unit 10 notifies the user of the recognition result.

[0022]

As described above, according to the voice recognition apparatus of the present invention, at the time of registration, the input voice at the time of registration is recognized by using the HMM parameter obtained by learning, and its likelihood information is obtained. By doing so, the rejection threshold can be set according to the user at the time of recognition, so there is no inconvenience that the user is rejected no matter how many times they say it, and the usability of the user is improved. Can be made.

[Brief description of drawings]

FIG. 1 is a configuration block diagram of a voice recognition device 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 when registering a voice recognition device according to an embodiment of the present invention.

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

FIG. 5: Conventional Hidden Markov Mode
Example of l

FIG. 6 is an example diagram showing correspondence between a voice waveform, a time series of feature data, and each state of the HMM in conventional voice recognition.

[Explanation of symbols]

 1 voice input means 2 word voice cutout unit 3 feature extraction unit 4 state number estimation unit 5 learning unit 6 likelihood output unit 7 voice dictionary file 8 threshold value setting unit 9 collation determination unit 10 determination result output unit 11 microphone 12 ROM 13 CPU 14 RAM 15 Monitor 16 File device

Claims (1)

[Claims] 1. A voice input unit 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 voice including the word voice, and a feature for extracting feature data from the cut out word voice. From the feature data and the HMM parameters, an extraction unit, a state number estimation unit that estimates the number of states for the word speech when modeling with the HMM from the feature data, a learning unit that applies the feature data to the word model and obtains HMM parameters. Likelihood output unit for obtaining a likelihood, a voice dictionary file composed of learned HMM parameters and likelihood information, a threshold setting unit for setting a threshold for rejection, and a likelihood for each word model. And a determination result output unit that outputs a recognition result.