JP3605011B2 - Voice recognition method - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition method.
[0002]
[Prior art]
In speech recognition, there are those using DP matching and those using a probabilistic model (HMM; Hidden Markov Model). In word recognition using DP matching, a template is created for each word. In word recognition using the HMM, a template is created in smaller units.
[0003]
In word recognition using DP matching, as is well known, distance is used as an evaluation value for evaluating the degree of matching between an input voice and a template. The smaller the distance, the higher the matching degree is determined.
[0004]
In word recognition using the HMM, as is well known, likelihood is used as an evaluation value for evaluating the degree of matching between an input speech and a template. It is determined that the matching degree is higher as the likelihood is larger.
[0005]
By the way, when a voice is uttered in a noisy environment, a consonant or the like having a low power has its acoustic characteristics contaminated by noise, and the likelihood (in the case of using HMM) obtained as a result of matching with a template or The distance (when DP matching is used) has low reliability. Also, if the noise pattern at the time of template creation and the noise pattern at the time of recognition are different, the reliability and the distance or likelihood in the noise section will be low.
[0006]
Therefore, conventionally, a method of improving a recognition accuracy of a noise section by learning a silence model from an actual noise environment and adding a silence model to a portion other than the speech section of the template has been proposed. However, this method cannot improve the recognition accuracy of a voice section that is originally low in power and buried in noise.
[0007]
[Problems to be solved by the invention]
SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech recognition method capable of improving recognition accuracy when performing speech recognition on speech uttered in an environment having noise.
[0008]
[Means for Solving the Problems]
A first speech recognition method according to the present invention is a speech recognition method for selecting an optimal standard pattern by determining a matching degree between an input speech feature and a template based on an evaluation value. A first step of generating a noise feature, in a section whose power is smaller than a predetermined threshold value of the template, calculating a first evaluation value by comparing a speech feature of the section of the template with an input speech feature; A second step of calculating a second evaluation value by comparing the noise feature created in one step with the input speech feature, and a higher matching degree of the first evaluation value and the second evaluation value Is adopted as the evaluation value for the section in question.
[0010]
A second speech recognition method according to the present invention is a speech recognition method for selecting an optimal standard pattern by determining a matching degree between an input speech feature and a template based on an evaluation value. The first step of creating a noise feature, in a section in which the power is smaller than a predetermined threshold of the template and in a section in which the power is smaller than the predetermined threshold of the input speech feature, the speech feature and the input speech feature of the template in the corresponding section. And a second step of calculating a second evaluation value by comparing the noise feature created in the first step with the input speech feature, and a first evaluation value by comparing Of the evaluation value and the second evaluation value, the evaluation value with the higher matching degree is used as the evaluation value for the section in the third step. Characterized in that it comprises a.
[0011]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[0012]
[1] Description of the concept of the present invention The concept of the present invention will be described.
[0013]
FIG. 1A shows a template, and FIG. 1B shows an input voice feature. In FIG. 1A, Th indicates a preset threshold for determining an evaluation section. In FIGS. 1A and 1B, sections A and A ′ indicate noise sections, sections B and B ′ indicate voice sections, and sections C and C ′ indicate noise sections.
[0014]
Hereinafter, the speech recognition method will be described.
[0015]
First, a noise feature is created from the noise section A 'or C' of the input speech feature.
[0016]
For the sections U1 and U2 where the power is larger than the threshold value Th of the template, evaluation is performed by comparing the audio features of such portions of the template with the input audio features. That is, an evaluation value (likelihood or distance) is calculated.
[0017]
Two types of evaluation are performed for sections A, D1, D2, D3, and C whose power is smaller than the template threshold Th. First, evaluation is performed by comparing the audio features of such a portion of the template with the input audio features. Second, the evaluation is performed by comparing the noise feature created from the noise section A ′ or C ′ of the input speech feature with the input speech feature. Then, the first evaluation result and the second evaluation result are compared, and the evaluation value with the higher matching degree is used as the evaluation value of the section. That is, if the evaluation value is likelihood, the likelihood with the larger evaluation value is used as the evaluation value of the section, and if the evaluation value is distance, the distance with the smaller evaluation value is used as the evaluation value of the section. Use as a value.
[0018]
In a section in which the spectrum, which is a feature of speech, is completely masked from noise because the SNR value is small, the matching degree when the noise feature is used as the template feature is high. On the other hand, in a section in which the original voice feature remains, the matching degree when the voice feature of such a portion of the template is used as the template feature is high. Therefore, as in the above-described embodiment, two types of evaluations are performed for sections A, D1, D2, D3, and C whose power is smaller than the threshold value Th of the template, and the evaluation value with the higher matching degree is used. And the accuracy of voice recognition is improved.
[0019]
In the above embodiment, two types of evaluations are performed for the sections A, D1, D2, D3, and C whose power is smaller than the threshold value Th of the template. However, the threshold value Th ′ of the input voice feature (FIG. 1B) Two types of evaluations may be performed for sections A ′, D1 ′, D2 ′, D3 ′, and C ′ with lower power.
[0020]
In addition, two types of evaluation may be performed in a section in which the power is smaller than the threshold value Th of the template and in which the power is smaller than the threshold value Th ′ of the input voice feature.
[0021]
[2] Description of Embodiment When Applied to Speech Recognition Using HMM An embodiment when applied to speech recognition using an HMM will be described.
[0022]
FIG. 2A shows a template of “SAKAI”. The template includes an HMM state sequence and power for each state.
[0023]
When learning the template, as shown in FIG. 2B, the template is learned from the learning voice data. In this case, the learning may be performed in units of words or in units smaller than words, for example, in units of phonemes. Further, the power for each state is obtained by EM estimation in the same manner as other parameters (for example, ceptoram). Alternatively, the power for each state is obtained by aligning the learning data with the model obtained by learning using a model learned using other parameters (for example, ceptoram), and obtaining the state of the learning data corresponding to each state. The average power of each state may be calculated from the alignment portion to obtain the average power.
[0024]
FIG. 3 shows a template and input speech features.
[0025]
First, a noise feature is created from a noise section of an input speech feature.
[0026]
The threshold value Th of the template (FIG. 2) is set to 5 in this example. For the portion where the power of the template is large, that is, for the portion “A” in the template, the likelihood is calculated by comparing the audio feature of such portion of the template with the input audio feature.
[0027]
Two types of evaluations are performed on the portion of the template (FIG. 2) where the average power is smaller than the threshold value Th, that is, on the "silence", "S", and "K" portions in the template. First, the likelihood is obtained by comparing the speech feature of such a portion of the template with the input speech feature. Second, a likelihood is obtained by comparing a noise feature created from a noise section of the input speech feature with the input speech feature. Then, the first evaluation result and the second evaluation result are compared, and the larger likelihood is used as the likelihood of the section.
[0028]
Note that two types of evaluation may be performed on the portions “silence”, “S”, “A”, and “I” whose power is smaller than the threshold value Th ′ of the input voice feature (FIG. 3).
[0029]
In addition, two types of evaluation are performed on portions "silence" and "S" that are portions having power smaller than the threshold value Th of the template (FIG. 2) and have power smaller than the threshold value Th 'of the input voice feature (FIG. 3). You may do so.
[0030]
A speech recognition method for performing two types of evaluations on a portion having a power smaller than the threshold value Th of the template and a power smaller than the threshold value Th ′ of the input speech feature (referred to as the method of the present invention), and a conventional speech recognition method (conventional method) The following comparative experiment was conducted.
[0031]
That is, in a noise environment with an S / N ratio of 5 dB, a total of five men and women make utterances for 100 place name words, and these voices are subjected to voice recognition by the method of the present invention and the conventional method to perform voice recognition. The recognition rate was determined. While the recognition rate was 94.8% in the conventional method, the recognition rate was improved to 96.0% in the method of the present invention.
[0032]
【The invention's effect】
According to the present invention, it is possible to improve recognition accuracy when performing speech recognition on a speech uttered in an environment with noise.
[Brief description of the drawings]
FIG. 1 is a waveform diagram showing a template and input speech features.
FIG. 2 is a schematic diagram showing a template in speech recognition using an HMM.
FIG. 3 is a schematic diagram showing templates and input speech features in speech recognition using an HMM.

Claims (2)

In a voice recognition method for selecting an optimal standard pattern by determining a matching degree between an input voice feature and a template based on an evaluation value,
A first step of creating a noise feature from a noise section of the input speech feature,
In the section where the power is smaller than the predetermined threshold value of the template, the first evaluation value is calculated by comparing the speech feature of the template with the input speech feature, and the noise feature created in the first step is compared with the noise feature created in the first step. A second step of calculating a second evaluation value by comparing the input voice feature with the input voice feature, and, of the first evaluation value and the second evaluation value, an evaluation value having a higher matching degree is evaluated for the section. The third step to take as value,
A voice recognition method comprising: In a voice recognition method for selecting an optimal standard pattern by determining a matching degree between an input voice feature and a template based on an evaluation value,
A first step of creating a noise feature from a noise section of the input speech feature,
In a section in which the power is smaller than the predetermined threshold of the template and in which the power is smaller than the predetermined threshold of the input speech feature, the first evaluation is performed by comparing the speech feature in the section of the template with the input speech feature. A second step of calculating a second evaluation value by calculating a value and comparing the noise feature created in the first step with the input speech feature; and
A third step of adopting, from the first evaluation value and the second evaluation value, the evaluation value with the higher matching degree as the evaluation value for the section;
A voice recognition method comprising:

Images