JPH0411297A - Voice recognition device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and more particularly to a configuration of a speech recognition device that improves speech recognition accuracy.

[Prior Art] In a conventional speech recognition device, input speech is preprocessed by amplifying it with an amplifier and passing through a frequency filter, and then is converted into an analog-to-digital signal and inputted to a speech recognition section.

The speech recognition unit converts the digitized speech into a time series of feature vectors, vector quantizes it, and converts it into a sequence of labels of reference vectors.

This label string is pattern matched with short sound data and recognized as speech.

[Problems and objects to be solved by the invention] In conventional methods, when the power value of the voice is used as part of the feature vector for recognizing the voice, the power value of the input voice is different from that of the reference vector. In this case, there is a problem in that the power value as a feature parameter significantly impairs recognition accuracy.

For example, even a slight distance between the microphone that inputs the voice and the mouth will change the voice recognition accuracy, making it impossible to always obtain the highest voice recognition accuracy.

An object of the present invention is to eliminate such drawbacks, use the power value of input speech as a feature parameter, improve speech recognition accuracy, and provide a speech recognition device that can enjoy stable speech recognition accuracy. It is in.

[Means for Solving the Problems] A speech recognition device of the present invention includes a conversion means for converting input speech into analog-to-digital data, a multiplication means for multiplying the output of the conversion means by a multiplier, and an output of the multiplication means, a feature extraction means for converting the power value of the input voice into a time series of feature vectors that include a part thereof; a vector quantization means for converting the previous feature vector into a label of a reference vector that is close in distance; an error detection means that calculates and time-averages the difference between the power value of the previous reference vector associated with the feature vector; and a multiplier of the multiplication means so that the output of the error detection means becomes zero. and a multiplier control means.

[Operation] According to the configuration of the present invention, the power value of the human voice is such that the power value as a component of the voice feature vector is on average equal to the power value as a component of the reference vector with which the feature vector is associated. Since feedback is applied and the volume of the voice is adjusted so that they match, the reliability of the power value as a feature parameter is improved, and recognition accuracy is improved and stabilized.

[Embodiment] FIG. 1 is a block diagram showing the configuration of an embodiment of the speech recognition device of the present invention. A speaker's voice is input through a microphone 1 and amplified by an amplifier 2.

The output of the amplifier 2 passes through a frequency filter 3, is sampled every 50 microseconds by an analog-to-digital converter 4, and is converted into a 12-bit digital audio signal.

The digitized audio waveform output from the analog-to-digital converter 4 is input to a multiplier 5 and multiplied by a multiplier.

The feature extraction means 6 extracts frames every 5 milliseconds from the speech waveform output from the multiplication means, with a time interval of 15 milliseconds as one frame, and calculates the L of the speech waveform of each frame.
Calculate PC cepstral coefficients up to the 12th order. The n-th LPC cepstral coefficient in the frame at time t is Cn
Write (t).

In addition, the sum of squares of the audio amplitudes within the same frame is calculated, and its logarithm is calculated, which is used as the power value for that frame.

Write this as w (t). Furthermore, the difference between the LPC cepstrum coefficients is calculated as follows.

Dn (t)=Cn (t+δ)-Cn(t-δ)n=
1.2, ・ ・ ・ ・ 12 Finally, take the difference in power values as follows.

x (t) = w (10 δ) −w (− δ) In both cases, δ is 15 milliseconds.

The above parameters CI (t), C2(t), CI
2 (t) , DI (t) , D2 (t)1
, DI2 (t) , w (t) x (t)
Create a 26-dimensional feature vector with . In this way, the feature extraction means takes a frame every 5 milliseconds,
Feature vectors are calculated to create a time series.

The codebook 7 consists of reference vectors for vector quantization. The reference vector is created in advance using learning data. A large amount of training utterance data is processed in the same manner as described above, a large amount of feature vectors are created, and the feature vector space is Lind based on the distribution of these.
It is divided into 256 separate spaces using the e-Buzo-Gray algorithm. Linde-Buzo-G
ray algorithm [Y, Linde et al.
n Algorithm for Vector Qu
antizerDesign” Positive EE Transac
tions on CommunicationCO
M-28(1), p84. January, 198
I am familiar with 0J. Vectors located at the center of each divided space are employed as reference vectors (256), and appropriate labels are attached to them.

The vector quantization means 8 calculates the distance between the feature vector that is the output of the feature extraction means and the reference vector using the following formula. The subscript r indicates a component of a reference vector. Also, the gradient indicates a square operation.

Σ (Ci-Cir) 82 + K ・Σ (D
i-D ir) 82i=1
i=1+L・(w-wr)
82 + M ・ (x-xr)
82 Here, L and M are the weights of each component in defining the distance between vectors, and in this example, they are each 0.7.
5.0.02.0.05. This distance calculation is performed for all reference vectors, the closest reference vector is selected, and its label is used as the label of this feature vector.

The short sound collation means 9 is r K-F, Lee et al., The
5PHINXSpeech Recongnition
System, ICASSP89. p445.1
989'', the short sound data modeled using the hidden Markov model is compared with the label string after vector quantization to find the most probable short sound.

As a result of matching using the short sound matching method, the likelihood of the correspondence is 85%.
If it exceeds , the difference between the power values of the associated feature vector and the reference vector is accumulated in the error detection means 10.

If it is recognized as a non-speech section, errors are not accumulated.

The error detection means 10 averages the cumulative value with a time constant of approximately 10 seconds, and transmits the cumulative error to the multiplier control means 11. The multiplier of the multiplier 5 is changed so that the cumulative error of the multiplier control means 11 becomes zero.

The power value of speech is an important feature of speech, and using power as a distance measure for vector quantization can improve recognition accuracy. Furthermore, since the multiplier of the multiplication means is controlled so that the power value matches the reference data itself, the accuracy of vector quantization is greatly improved, and as a result, the accuracy of speech recognition is improved.

[Effects of the Invention] As described above, according to the present invention, speech recognition with high recognition accuracy becomes possible.

[Brief explanation of drawings]

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

Claims (1)

[Scope of Claims] Conversion means for converting input audio into analog-to-digital data; Multiplication means for multiplying the output of the conversion means by a multiplier; The output of the multiplication means includes a power value of the input audio as a part thereof. A feature extraction means for converting the feature vector into a time series; a vector quantization means for converting the previous feature vector into a label of a reference vector having the closest distance; The present invention is characterized by comprising: error detection means for calculating and time-averaging the difference in power values of the preceding reference vectors; and multiplier control means for changing the multiplier of the multiplication means so that the output of the error detection means becomes zero. voice recognition device.