JPH10198394A - Voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition method with which a voice superposed with noise can be recognized with a high degree of provability even though a kind of noise and an SN ratio are simultaneously changed on issuance of voice. SOLUTION: A voice recognition device comprises a voice signal input part 2, a characteristic value extracting part 2, a data storage part 3 and a recognition result determining part 4. Upon creation of a word HMM(hiddened Markov model) used for recognition, noise HMM having been created previously from several kinds of noises, is used, and further, in consideration with S/N rations in a plurality of levels, it is possible to recognize a voice with a high degree of provability even though a kind of noise to e superposed, and the S/N ratio will be changed on the way.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech recognition method using a hidden Markov model (HMM).

[0002]

2. Description of the Related Art An HMM method is widely used for automatic speech recognition by a computer. An HMM is generally a non-deterministic stochastic finite state automaton having a plurality of states, each of which is a stochastic stationary signal source. That is, the HMM can be said to be a non-stationary signal source that emits a signal while non-deterministically switching the stochastic stationary signal source.

There is a method called maximum likelihood estimation method as a method used for speech recognition using the HMM. In order to use the HMM for word recognition, first, an HMM is prepared for each word to be recognized, and the HMM is set so that each HMM can easily output a feature parameter sequence extracted from a speech sample belonging to the own word. Adjust the defined internal parameters. In the maximum likelihood estimation method, when an unknown voice is input, the output ease (likelihood) of a feature parameter sequence extracted from the unknown voice is calculated for each HMM, and the maximum likelihood is output. A word corresponding to the HMM is set as a recognition result.

[0004] One of the methods for recognizing a word voice on which noise is superimposed using an HMM using a word as a recognition unit is Tran.
c Martin wrote the document `` Recognition
ofNoisy Speech by Composi
Tion of Hidden Markov Mod
Els '' (NO. SP92-96)
There is a method using a VO-HMM. This is because the noise HMM and the internal parameters of the word HMM are synthesized by a method called NOVO conversion in the above-mentioned document, and the N
By using the OVO-HMM, word speech on which noise is superimposed is recognized with high accuracy.

FIG. 6 is a conceptual diagram of a conventional NOVO conversion. In the conventional method, a recognition target word HMM is generated by learning using learning sample data of the recognition target word,
After generating a noise HMM by learning using one type of noise learning sample data, these recognition target words HMM
And the noise HMM are combined by NOVO conversion to obtain a NOVO-HMM for each recognition target word.

[0006]

SUMMARY OF THE INVENTION Conventional NOVO-HM
To obtain a high recognition rate in the recognition method using M,
The noise used when creating the VO-HMM, ie,
It is necessary that the noise to be considered during recognition does not change significantly during the utterance time, and if the type of noise or the SN ratio changes significantly during utterance, there is a problem that the recognition rate is greatly reduced. .

Accordingly, an object of the present invention is to provide a speech recognition method capable of recognizing a noise-superimposed speech with a high probability even when the type of noise and the SN ratio simultaneously change during speech production.

[0008]

According to the first aspect of the present invention,
A speech recognition method for recognizing speech on which noise is superimposed using an HMM having a word as a recognition unit, wherein a noise HMM previously generated from a plurality of types of noise is used when generating a word HMM used for recognition. In addition, by considering the SN ratios of a plurality of levels, even when the type of noise to be superimposed and the SN ratio change on the way, the voice is recognized with a high probability.

According to a second aspect of the present invention, there is provided a speech recognition method for recognizing a speech on which noise is superimposed using an HMM having a word as a recognition unit, wherein a plurality of noises and a plurality of levels of SN ratio are considered. When calculating the likelihood of each generated word HMM, the type of noise on the path that gives the maximum likelihood is recorded, and the transition status of the type of noise in the first few words HMM is almost the same. If so, the transition sequence of the noise is fixed to the previous transition state, the likelihood of each word HMM is calculated, and the calculation amount is reduced.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention considers a plurality of types of noise and a plurality of levels of S / N ratios in generating a NOVO-HMM that recognizes a word voice on which noise is superimposed, so that the type of noise or voice NOVO- does not significantly reduce recognition accuracy even if the SN ratio of noise and noise changes during the time of occurrence.
An HMM is generated.

Hereinafter, a speech recognition method according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration block diagram of a speech recognition device according to an embodiment of the present invention. 1 is an audio signal input unit for converting an audio signal, which is learning sample data or recognition target data, into a digital value, 2 is a feature amount extraction unit for calculating a feature amount for each frame from the input signal, 3 is learning sample data, noise data , A data storage unit for storing a recognition target word HMM, a noise HMM, and a NOVO-HMM, and a recognition result determination unit 4 for calculating an output probability of an input word and determining a recognition result.

FIG. 2 is a circuit block diagram of a speech recognition apparatus according to one embodiment of the present invention. 11 is a microphone, 12 is a central processing unit (CPU), 13 is a read-only memory (ROM), 14 is a writable memory (RAM), 1
5 is an output device.

The audio signal input device 1 shown in FIG. 1 comprises a microphone 11 and a CPU 12, and a feature extracting unit 2, a data storing unit 3, and a recognition result judging unit 4 have a CPU 12
13 is executed by executing a program written in the RAM 13 and accessing the RAM 14.

FIG. 3 is a flowchart of a voice recognition method according to an embodiment of the present invention. First, a voice is input from the microphone 11, and the feature amount obtained through the feature amount extraction unit 2 is written into the RAM 14 (step 2). Next, ROM1
No. 3 created for each recognition target word stored on
For the VO-HMM, the likelihood for the feature on the RAM 14 is calculated (step 3). Then, a word name corresponding to the NOVO-HMM that has output the maximum likelihood is output to the output device 15 as a recognition result (step 4).

FIG. 4 is a conceptual diagram of NOVO conversion used in the speech recognition method according to one embodiment of the present invention.
This figure shows the process of creating an OVO-HMM,
The number of states of the noise HMM to be transformed is 2, the noises to be considered at the time of recognition are A and B, and the number of levels of the SN ratio is x [d
B] and y [dB].

The shape of the noise HMM generated by the learning using the learning sample data differs from the conventional method of FIG. In the conventional method, a two-state noise HMM is directly generated from one type of noise by learning, and then the noise HMM is generated.
NOVO conversion is performed on the word HMM and the word HMM. However, in the present invention, first, a one-state noise HMM is generated by learning from two types of noise, and thereafter, the number of noise HMMs of each noise HMM is increased by the number of SN ratio levels to be considered. Duplicate state. Then, the state transition probabilities between the states of all the noise HMMs obtained by learning and duplication are artificially given (in this case, the self transition probability is about 0.7, and the other state transition probabilities are about 0.1, respectively). ) Combine the states of the noise HMM to generate a 4-state noise HMM. When the conventional NOVO conversion is performed on the four-state noise HMM and the recognition target word HMM, the shape of the NOVO-HMM is such that four rows of word HMMs are arranged as shown at the bottom of FIG. It takes shape. Each state in the top row uses a coefficient for x [dB] of noise A at the stage of adding the speech linear spectrum and the noise linear spectrum in the NOVO conversion. A using the coefficient for y [dB] of A, each state in the third column is x [d
The state using the coefficient for B] and the state in the fourth column use the coefficient for y [dB] of the noise B.

FIG. 5 is a conceptual diagram of the calculation amount reduction of the voice recognition method according to one embodiment of the present invention. The NOVO-HMM created by the above-described method has a NOVO-HMM when only a single type of noise or only a single-level SN ratio is considered.
It has more states than HMM, and each NOVO-
An enormous amount of calculation is required for the likelihood calculation of the HMM. Therefore, when calculating the likelihood of each NOVO-HMM, the type of noise on the path that gives the maximum likelihood and the SN ratio are recorded, and the transition state of the noise and SN ratio in the NOVO-HMM for the first few words is recorded. When they are almost the same, the sequence of the transition between the noise and the S / N ratio is fixed, and the likelihood of all NOVO-HMMs is recalculated. As a result, the amount of calculation can be significantly reduced.

[0018]

According to the speech recognition method of the present invention, noise-superimposed speech can be recognized with a high probability even when the type of noise and the S / N ratio change simultaneously during speech production.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a speech recognition device according to an embodiment of the present invention;

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

FIG. 3 is a flowchart of a voice recognition method according to an embodiment of the present invention;

FIG. 4 is a conceptual diagram of NOVO conversion used in a speech recognition method according to an embodiment of the present invention.

FIG. 5 is a conceptual diagram of a calculation amount reduction of the voice recognition method according to one embodiment of the present invention;

FIG. 6 is a conceptual diagram of a conventional NOVO conversion.

[Explanation of symbols]

 Reference Signs List 1 audio signal input unit 2 feature amount extraction unit 3 data storage unit 4 recognition result determination unit 11 microphone 12 CPU 13 ROM 14 RAM 15 output device

Claims (2)

[Claims] 1. A speech recognition method for recognizing speech on which noise is superimposed using an HMM having a word as a recognition unit, wherein a word HMM used for recognition is generated in advance from a plurality of types of noise. Speech recognition characterized by using a noisy noise HMM and considering a plurality of levels of S / N ratios to recognize speech with high probability even when the type of superimposed noise and the S / N ratio change on the way. Method. 2. A speech recognition method for recognizing speech on which noise is superimposed by using an HMM having a word as a recognition unit, wherein each of the word HMMs generated in consideration of a plurality of noises and an S / N ratio of a plurality of levels. When calculating the likelihood, the type of noise on the path that gives the maximum likelihood is recorded, and the first few words HM
When the transition state of the noise type is substantially the same in M, the likelihood of each word HMM is calculated by fixing the sequence of the noise transition to the previous transition state, and the amount of calculation is reduced. Voice recognition method to be used.