JPH096381A - Voice word recognition method - Google Patents

Abstract

PURPOSE: To make it possible to carry out recognition of silence adapted to the environment without re-learning even when the voice input environment is changed by performing a phoneme recognition processing forcedly allocating silence data to frame in a detected silence section. CONSTITUTION: A silence section detection part obtains logarithmic power at every frame of an input voice by using the power information of the voice, and detects a vice section that the logarithmic power becomes a threshold or below as the silence section. Inputted feature vector time sequence is converted into phoneme label time sequence by using a phoneme identifier by a neural net. In such a case, the silence is recognized as one phoneme, too. At this item, when the input environment when voice recognition is really performed is different from the same of the time of learning, since voice waveform is affected by a disturbance such as a noise of a microphone, etc., the silence data are allocated forcedly to the frame in the silence section obtained by the silence section detection part for the voice labeled phoneme recognized result. Even when the silence is not recognized as the silence, no influence is affected to a word recognition part thereafter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech word recognition method which eliminates the environment dependence of silent parts.

[0002]

2. Description of the Related Art The concept of current speech recognition technology is shown in FIG. When performing word recognition, a speech waveform is sampled at a certain time interval, converted into a time series of multidimensional feature vectors such as a spectrum, and then handled. Here, the feature vector is associated with some phoneme label. In this case, silence is also treated as one phoneme. Similarly, the words to be recognized are converted into time series of multidimensional feature vectors, and these are registered in the computer as standard patterns. Again, silence is treated as one phoneme for words where silence appears in the word.

In the voice recognition process, the similarity between the input feature vector time series and the standard pattern feature vector time series is calculated for all standard patterns, and the word of the most similar standard pattern is used as the recognition word.

However, in general, it is not possible to directly compare the input feature vector time series and the standard pattern feature vector time series as they are. This is because there are individual differences in the length of time that a person utters a sentence or word, and even if the same person utters the same word, it varies greatly depending on the mood depending on the day. Moreover, at this time, the expansion and contraction of the vocalization time is not uniform, but varies nonlinearly.

In the DP matching method, the time axis is converted by using the dynamic programming method so that the feature vector time series of the input speech best matches the feature vector time series of the standard pattern, and then the similarity is calculated. Ask.

The concept of this DP matching is shown in FIG. In the figure, the horizontal axis represents the input voice and the vertical axis represents the standard pattern of words registered in the computer in advance. Here, it is assumed that both the input voice and the standard pattern are described in time series of phoneme labels, not in time series of feature vectors.

[0007]

In the conventional speech recognition method, silence can be learned as one phoneme. In this case, for example, like 1 minute
When recognizing a word in which a silent portion always exists between “i” and “pu”, a template considering silence is created in advance such as i-pun. However, since the input voice data is converted into digital data through the process of A / D conversion from the actual voice through a microphone, especially for a silent portion, the input method at the time of input capturing Or, the waveform is usually different due to the difference in the characteristics of the capturing device.

Therefore, when the input environment of the phoneme data used at the time of learning is different from the voice input environment in which voice recognition is actually used, the silent part may be erroneously recognized as another phoneme.

Therefore, in order to correctly recognize the silence in the neural network parameters after learning in another environment, the input environment exactly the same as that at the time of learning is prepared again, or conversely, re-learning of the silent part is performed in the system. There is nothing else you can do for it, and it will be very troublesome.

The present invention has been made in view of the above points, and an object thereof is to provide a speech word recognition method capable of recognizing silence without adapting to the environment even if the speech input environment changes. To do.

[0011]

According to the present invention, (1) the logarithmic power P (i) of each frame of input speech is

[0012]

[Equation 2]

And the logarithmic power P
(I) Silence section detection processing in which a speech section in which a threshold value θ is equal to or less than a predetermined threshold value is a silent section, and feature extraction is performed by discrete Fourier transform for each frame, and an input feature vector time series is set at a phoneme label time. A phoneme recognition process of recognizing a phoneme by converting it into a sequence, and forcibly allocating silence data to a frame within a silence interval detected by the silence interval detection process of the phoneme recognition result, and the phoneme recognition The word recognition that obtains the similarity between the phoneme label data obtained by the processing and the standard template of the phoneme label time series held as a dictionary by the DP matching method, and outputs the result of the standard template with the closest similarity as the speech recognition result. (2) The predetermined threshold value θ is set as follows: the phoneme data of a phoneme with low voice power and the phoneme data of a silent state. Are recorded, the frame power is calculated for each of the audio data, and the average value of the calculated frame powers is set to the center value. (3) For each frame of the input audio In addition, the feature extraction is performed by the discrete Fourier transform, and the phoneme recognition process that converts the input feature vector time series into the phoneme label time series is performed. The phoneme label data obtained by the phoneme recognition process and the similarity to the standard template are set to D by replacing the phoneme with such phonemes.
And a word recognition process of outputting the result of the standard template having the closest similarity degree as a speech recognition result by the P matching method. (4) The phoneme in which the silence is likely to be erroneously recognized is in a silent state. Of the phoneme data is recorded, the feature extraction is performed on the voice data by the discrete Fourier transform, the input feature vector time series is converted into a phoneme label time series, and the most phoneme result among the converted phoneme results is obtained. It is characterized by that.

[0014]

In the phoneme recognition process according to the first and second aspects of the invention, since the silence data is forcibly assigned to the frames within the detected silence section, the silence which is easily affected by noise is regarded as silence. Even if no phoneme is recognized, silence is not recognized as a wrong word.

In the word recognition processing of the invention of claim 3, the silent part in the standard template is set by replacing it with a phoneme in which silence is likely to be erroneously recognized, and in the word recognition processing of the invention of claim 4, Since the silent part in the standard template is set by replacing it with the most phoneme of the phoneme results when the data was recorded in the silent state, the phoneme of the silent part of the input speech to be recognized is set by the above replacement. It is extremely likely that the phoneme most closely resembles the phoneme that has been identified, and false recognition of words is avoided.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 1 shows a flow chart of the method of this embodiment. The input voice (step S ₁ ) is first input to the silent section detector (step S ₂ ), and the silent section of the voice uttered here is detected. Next, the phoneme recognition is performed in the phoneme recognition section (Step S _3). And this labeled data is performed word recognition passed to word recognition unit (Step S _4), the recognition result is obtained (Step S _5).

The silent section detecting section, the phoneme recognizing section and the word recognizing section will be described in detail below. (1) Silence section detection unit Detects a silence section from voice data using voice power information. (1-1) The logarithmic power of each frame of the input voice is calculated by the following mathematical expression.

[0018]

(Equation 3)

Here, P (i): logarithmic power of the i-th frame AM _K : amplitude of the k-th speech in the section included in the i-th frame.

(1-2) Then, the logarithmic power is the threshold value θ.
The following voice section is detected as a silent section as shown in FIG. Here, depending on the value of the threshold value θ, the phoneme (s),
In the fricative sounds such as (f) and (h), since the voice power is low, this phoneme section may be a silent section. Therefore, it is assumed here that the threshold value θ is an optimum value capable of distinguishing the silent section and the voiced section.

(1-3) The threshold value θ is obtained according to the flowchart of FIG. First, in an actual recognition system, phonemes (s) and (h) with low voice power are used.
A plurality of voices including is voiced to record voice data (step S ₁ ). Make sure that the captured data range does not include silence. Specifically, a recording start signal and a recording end signal are issued, and during this period, the voice is continuously uttered. Similarly, the state where no voice is input (silent state)
In, to record the voice data (step S _3). Then based on the data obtained in step S _1, S _3, by a method using each of the formulas (1) to determine the frame power in each of the voice section (Step S _2, S _4). Then, the average value of the frame power is obtained (step S ₅ ). Then, the central value of the respective frame power average values obtained as described above is set as the threshold value θ (step S
₆ ).

(2) Phoneme Recognition Unit Phoneme recognition of voice data is performed and phoneme labeling is performed. (2-1) For each frame, DFT (Discert)
Feature Transform (discrete Fourier transform) is performed to obtain a feature label time series.

(2-2) The input feature vector time series is converted into a phoneme label time series by using a phoneme classifier based on a neural network. At this time, silence is also recognized as one phoneme. At this time, if the voice input environment used when learning the neural network and the input environment when actually performing voice recognition by using the learned results are different, A disturbance such as noise from a microphone may affect the voice waveform until voice data is input to the section. Especially, the voice waveform of a silent portion is easily affected. In this case, silence may not be recognized as silence and may be erroneously recognized by another phoneme.

(2-3) Therefore, the present invention solves this problem by the following method. With respect to the phoneme recognition result labeled with the voice, the silence data is forcibly assigned to the frame of the silence section obtained by the silence section detection unit as shown in FIG. As a result, even if silence is not recognized as silence, it does not affect the subsequent word recognition unit. Above (1-2),
By setting the threshold value θ in (1-3) to an optimum value for the system, the environmental dependency of the silent section is eliminated.

(3) Word Recognition Unit The similarity between the phoneme label data obtained by the phoneme recognition unit and all standard templates prepared in advance is obtained by DP matching, and the result of the closest standard template is recognized by speech. Output as a result.

(Embodiment 2) FIG. 5 shows a flow chart of the method of this embodiment. Input voice (step S ₁ )
First, phoneme recognition is performed by the phoneme recognition unit (step S ₂ ). And this labeled data is performed word recognition passed to word recognition unit (Step S _3), the recognition result is obtained (Step S _4). Each of the phoneme recognition unit and the word recognition unit will be described in detail below.

(1) Phoneme Recognition Unit Phoneme recognition of voice data is performed and phoneme labeling is performed. (1-1) DFT (Discert) is performed for each frame.
Feature Transform (discrete Fourier transform) is performed to obtain a feature label time series.

(1-2) The input feature vector time series is converted into a phoneme label time series using a phoneme classifier based on a neural network. At this time, silence is also recognized as one phoneme.

(2) Word recognition section The degree of similarity between phoneme label data obtained by one phoneme and all standard templates prepared in advance is obtained, and the result of the closest standard template is output as a speech recognition result. .

(2-1) The silent part in the standard template is replaced with a phoneme in which silence is likely to be erroneously recognized. As a result, silence does not have to be recognized as silence due to the difference in the voice input environment.

A method of setting a phoneme in which silence is likely to be erroneously recognized by the word recognition section will be described below with reference to the flowchart of FIG. First, in a system that actually recognizes voice, a state where no voice is input for a certain period of time (silence state)
Then, the voice data is recorded (step S ₁ ). Next, phoneme labeling is performed using a phoneme classifier based on a neural network (step S ₂ ). Then, based on the phoneme results labeled with phonemes in step S ₂ , the most phoneme recognition result is set as a phoneme in which silence is likely to be erroneously recognized (steps S ₃ and S ₄ ).

(2-2) The similarity between the phoneme label data obtained by the phoneme recognition unit and all standard templates prepared in advance is obtained by DP matching, and the result of the closest standard template is the speech recognition result. Is output (step S _{4 in} FIG. 5).

[0033]

As described above, according to the present invention, the following excellent effects can be obtained. (1) According to the invention described in claims 1 to 4, when the voice input environment is changed, it is possible to recognize the silence, which is adapted to the environment, without re-learning.

(2) In the phoneme recognition processing according to the first and second aspects of the present invention, since the silence data is forcibly assigned to the frames within the detected silence section, the silence that is easily affected by noise is generated. , Even if the phoneme is not recognized as silence, the silence is not recognized as a wrong word.

(3) In the word recognition process of the invention of claim 3, the silent part in the standard template is set by replacing it with a phoneme which is likely to cause false recognition of the silence, and the word recognition process of the invention of claim 4 In the above, since the silent part in the standard template is set by replacing it with the most phoneme of the phoneme results when recording the data, silence does not have to be recognized as silence due to the difference in the voice input environment. become. Moreover, since the phoneme of the silent part of the input voice is very likely to be most similar to the phoneme set by the replacement, erroneous recognition of a word can be avoided.

[Brief description of drawings]

FIG. 1 is a flowchart showing an embodiment of the present invention.

FIG. 2 is a graph for explaining the operation of a silent section detection unit in one embodiment.

FIG. 3 is a flowchart showing a method of setting a threshold value θ in one embodiment.

FIG. 4 is an explanatory diagram showing an example in which a silent section is applied to a recognition result in one embodiment.

FIG. 5 is a flowchart of another embodiment of the present invention.

FIG. 6 is a flowchart illustrating a main part of another embodiment.

FIG. 7 is a conceptual diagram of a voice recognition method.

FIG. 8 is a graph illustrating a state of a DP matching method.

Claims (4)

[Claims] 1. A logarithmic power P for each frame of input speech.
Let (i) be the following equation, P (i): Logarithmic power of the i-th frame AM _K : Speech segment in which the logarithmic power P (i) is less than or equal to a predetermined threshold value θ obtained by the amplitude of the k-th speech in the segment included in the i-th frame Silence interval detection process and a feature extraction by discrete Fourier transform for each frame, to recognize the phoneme by converting the input feature vector time series into a phoneme label time series, the phoneme recognition result Of these, phoneme recognition processing forcibly allocating silent data to frames in the silent section detected by the silent section detection processing, and phoneme label data obtained by the phoneme recognition processing,
A speech characterizing process for obtaining a similarity between a phoneme label time series standard template held as a dictionary and a standard template by a DP matching method and outputting a result of the standard template having the closest similarity as a speech recognition result. Word recognition method. 2. The predetermined threshold value θ is obtained by recording voice data of phonemes with low voice power and phoneme data in a silent state, calculating the frame power of each of the voice data, and calculating the frame power. The speech word recognition method according to claim 1, wherein the frame power is set to a central value of average values of the frame powers. 3. A phoneme recognition process for converting an input feature vector time series into a phoneme label time series by performing feature extraction by a discrete Fourier transform for each frame of input speech, and a standard template of a phoneme label time series held as a dictionary. The silent part in the inside is set by replacing it with a phoneme in which silence is likely to be erroneously recognized, and the similarity between the phoneme label data obtained by the phoneme recognition process and the standard template is obtained by the DP matching method, and the similarity is calculated. And a word recognition process for outputting the result of the closest standard template as the speech recognition result. 4. A phoneme whose silence is likely to be erroneously recognized is recorded as voice data in a silent state, and feature extraction is performed on the voice data by discrete Fourier transform to convert the input feature vector time series into a phoneme label time series. The speech word recognition method according to claim 3, wherein the phoneme result is converted and is the most phoneme result among the converted phoneme results.