JPH04121794A - Speech recognizing method - Google Patents

Abstract

PURPOSE:To obtain tolerance to spectrum distortion and a high recognition rate by removing parameters which are smaller than a threshold value and then finding difference values in feature parameter between frames, generating a time-series pattern of the difference values, and calculating the similarity between each speech and a standard pattern on a statistical distance scale. CONSTITUTION:1. A speech sample is inputted to a speech input part 1. 2. A feature extraction part 12 obtains frequency characteristics, frame by frame, through a band- pass filter. 3. A power decision part compares effective values of power of the frequency characteristics of the respective frames with the threshold value theta and removes the feature parameters which are smaller than the threshold value theta. 4. A difference value generation part 14 finds the difference values between frames and a time series pattern generation part 15 generates the time series pattern of the difference values. 5. A similarity calculation part 17 calculates the similarity between the time series pattern of the difference values and standard patterns of respective speeches stored in a dictionary part 16 on the statistical distance scale. 6. A decision part 18 decides the recognition result which has the highest similarity. Consequently, speech recognition wherein the influence of distortion and noises is removed is enabled regardless of whether the power is large or small.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition method suitable for recognizing words from input speech using online terminals such as electric locks and IC cards.

[Prior Art] Conventionally, a speech recognition method as described in Japanese Patent Application Laid-Open No. 1-260490 has been proposed. This speech recognition method calculates the feature parameters of input speech in units of frames of a predetermined length, determines the difference value of the feature parameters within the frame, creates a time series pattern of the difference value, and creates a time series pattern of the difference value. This method calculates the similarity between the standard pattern of each voice and the standard pattern of each voice using a statistical distance measure, and performs voice recognition.

[Problems to be Solved by the Invention] In the prior art, speech recognition is performed by directly using the feature parameters of all frames of input speech, regardless of the power level of those frames.

However, since information on frames with low power is easily affected by distortion and stationary noise in the transmission system, the reliability of similarity determination is low.

Furthermore, since the difference value between frames with low power is determined by taking the difference in the frequency domain, it is treated in the same way as the difference value between frames with high power, and this has a large effect on the recognition rate.

That is, in the prior art, it is difficult to ensure a high recognition rate because information of a small frame with low power and low reliability of similarity determination has a large influence on the recognition rate.

An object of the present invention is to provide a speech recognition method that is resistant to stationary spectral distortion and can ensure a high recognition rate.

[Means for solving the problem] The present invention according to claim 1 calculates the characteristic parameters of input speech in units of frames of a predetermined length, and calculates whether the effective value of the power of each frame is smaller than an arbitrary threshold value. After excluding the feature parameters of the frame, calculate the difference value of the feature parameters between frames, create a time series pattern of the difference value, and compare the time series pattern of this difference value with the standard pattern of each voice. This method calculates similarity using a statistical distance measure and performs speech recognition.

The present invention as set forth in claim 2 calculates the characteristic parameters of input audio in units of frames of a predetermined length, and when the effective value of the power of each frame is smaller than an arbitrary threshold value, the influence of the characteristic parameters of the frame is calculated. After weighting is performed to reduce the difference between frames, the difference value of the feature parameters between frames is calculated, a time series pattern of the difference value is created, and the similarity between the time series pattern of this difference value and the standard pattern of each voice is calculated. is calculated using a statistical distance measure to perform speech recognition.

[Operation] According to the present invention, feature parameters of small frames that are easily affected by distortion and stationary noise in the transmission system and have low power and low reliability for similarity determination are excluded or their influence is reduced. After weighting, difference values of feature parameters between frames are determined, and speech recognition is performed based on this difference value.

That is, compared to the method that uses difference values between all frames regardless of the power level, speech recognition is performed while eliminating the effects of transmission system distortion and stationary noise. Therefore, it is possible to obtain a speech recognition method that is resistant to stationary spectral distortion and can ensure a high recognition rate.

[Embodiment] FIG. 1 is a schematic diagram showing a speech recognition system according to an embodiment of the present invention.

The speech recognition system 10 includes a speech input section 11, a feature extraction section 12, a power determination section 13, a difference value creation section 14, a time series pattern creation section 15, and a dictionary section (standard pattern storage section) 16.
, a similarity calculation section 17, and a determination section 18.

Hereinafter, a dictionary creation procedure and a recognition procedure using the speech recognition system 10 will be explained.

(A) A voice sample is taken into the voice input section 11.

At this time, the recognized words were the names of 47 prefectures, and the specific speaker was one person.

TB) Dictionary creation ■The known input speech waveform of each recognized word is passed through a 16-channel bandpass filter in the feature extraction unit 12,
Frequency characteristics are obtained every frame (12.8 m5ec).

■The power establishing unit 13 compares the experimentally determined threshold value θ with the effective value of the power of the frequency characteristic of each frame, and excludes the characteristic parameters of frames whose effective value of power is smaller than the threshold value θ. .

■The difference value creation unit 14 calculates the difference value of the feature parameters between frames, and the time series pattern creation unit 15 creates a time series pattern of the difference value. The time series pattern is stored in the dictionary section 16 and used as a dictionary.

(C) Recognition ■The unknown input speech waveform of each recognition word with stationary noise added is passed through a 16-channel band-pass filter in the feature extraction unit 12 for one frame (12,8m5e
c) Obtain frequency characteristics for each step.

■The power determination unit 13 compares the effective value of the power of the frequency characteristic of each frame with the experimentally determined threshold θ, and excludes characteristic parameters of frames whose effective value of power is smaller than the threshold θ. .

(2) A difference value creation unit 14 calculates a difference value between frames, and a time series pattern creation unit 15 creates a time series pattern of the difference values.

(2) The similarity calculation unit 17 calculates the similarity between the time series pattern of the difference values created in (2) above and the standard pattern of each voice stored in the dictionary unit 16 using a statistical distance measure.

(2) In the determination unit 18, the one with the highest degree of similarity as a result of (2) above is determined as the recognition result.

However, experimental results of the conventional method and the method of the present invention using the speech recognition system 10 described above will be explained.

(Conventional method) Experiment: When calculated using a statistical distance measure using inter-frame difference values of feature parameters (outputs of band-pass filters).

The specific speaker was one person, and the recognized words were the names of 47 prefectures.

Results: The recognition rate was 93.2%.

(Method of the present invention) Experiment: When feature parameters with small power (output of band pass filter) are excluded, inter-frame difference values are used as input, and recognition is performed using a statistical distance measure.

The specific speaker was one person, and the recognized words were the names of 47 prefectures.

Results: The recognition rate was 95.3%.

In the implementation of the present invention, the dictionary creation stage and the recognition stage 1i1 (stage (B) (■) and (C) (■) described above) do not exclude feature parameters of frames with low power. , weighting may be performed to reduce the influence of the characteristic parameters of the frame.

The voice recognition system 10 has the following effects.

According to the above embodiment, feature parameters of small frames that are easily affected by transmission system distortion and stationary noise and have low power or reliability for similarity determination are excluded or weighted to reduce the influence. After that, difference values of feature parameters between frames are determined, and speech recognition is performed based on this difference value.

That is, compared to the method that uses difference values between all frames regardless of the power level, speech recognition is performed while eliminating the effects of distortion and stationary noise in the transmission system. Therefore, it is possible to obtain a speech recognition method that is resistant to stationary spectral distortion and can ensure a high recognition rate.

[Effects of the Invention] As described above, according to the present invention, it is possible to obtain a speech recognition method that is resistant to stationary spectral distortion and can ensure a high recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram showing a speech recognition system according to an embodiment of the present invention. DESCRIPTION OF SYMBOLS 10... Speech recognition system, 11... Speech input unit, 12... Feature extraction unit, 13... Power determination unit, 14... Difference value creation unit, 15... Time series pattern creation unit , 16... Dictionary section, 17... Similarity calculation section, 18... Judgment section. Patent applicant: Sekisui Chemical Co., Ltd. Representative Hiroshi 1) Kaoru

Claims (2)

[Claims] (1) Calculate the feature parameters of the input audio in units of frames of a predetermined length, and when the effective value of the power of each frame is smaller than an arbitrary threshold, after excluding the feature parameters of the frame, Speech recognition that calculates the difference values of parameters, creates a time series pattern of the difference values, calculates the similarity between this time series pattern of the difference values and the standard pattern of each voice using a statistical distance scale, and performs speech recognition. Method. (2) Calculate the feature parameters of the input audio in units of frames of a predetermined length, and when the effective value of the power of each frame is smaller than an arbitrary threshold, weighting is applied so that the influence of the feature parameters of the frame is reduced. After that, calculate the difference value of the feature parameters between frames, create a time series pattern of the difference value, and calculate the similarity between this time series pattern of the difference value and the standard pattern of each voice using a statistical distance measure. A voice recognition method that performs voice recognition.