JPH09331391A - Speech quality object estimate device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the speech quality estimate accuracy by integrating a section to identify a coding system of a voice signal so as to select a standard deteriorating voice signal and a weight coefficient in matching with the coding system of a test voice signal. SOLUTION: A coding voice identification section 1 identifies a coding system of a coding voice signal having a code error. A similarity calculation section 3 obtains a time series of the similarity between the analysis result by a voice analysis section 2 and a standard deterioration voice signal selected based on the identification result by the coding voice identification section 1. A weight coefficient database 6 stores weight coefficients between units obtained by subject evaluation values of various deteriorated voice signals and learning processing by a neural net by the similarity time series. A quality estimate calculation section 4 applies input processing to the weight coefficient selected based on the time series of the similarity obtained by the similarity calculation section 3 and the identification result of the coded voice identification section 1 from the weight coefficient database 6 and estimates the speech quality.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech quality objective estimation apparatus, and more particularly to a speech quality objectively estimating speech speech quality of a voice signal deteriorated by distortion and noise generated in a telephone transmission apparatus, by a physical quantity of the speech signal. The present invention relates to an objective estimation device.

[0002]

2. Description of the Related Art A conventional example will be described with reference to FIG. FIG.
In FIG. 1, the portion surrounded by the chain line shows a conventional example.
A speech quality objective estimation apparatus that objectively estimates speech quality using a pattern matching method and a neural network inputs a test speech signal, which is a speech signal to be quality-estimated, to a speech analysis unit 2 and outputs this analysis result. , The similarity calculation unit 3 performs pattern matching processing on the similarity with the standard deteriorated voice signal in which the characteristics of the deteriorated voice signal stored in the standard deteriorated voice database 5 in advance are represented by a short time series of parameters. Then, the time series of the obtained similarity is input to the quality estimation value calculation unit 4 forming the neural network to estimate the call quality. The calculation of the degree of similarity in the degree-of-similarity calculation unit 3 that performs the pattern matching of this estimation device and the learning process for creating the standard deteriorated speech signal are performed in the same manner as the speech quality objective measuring method (Japanese Patent Application No. 3-118924). Then, a MOS (Mean Opinion Sc) representing a measure of speech quality with the similarity time series as an input.
The ore: average opinion score) is output to the neural network, which is composed of three layers: an input layer, an intermediate layer, and an output layer. Learning processing and quality estimation processing are performed as follows.

In the learning process, first, the time series of the similarity between the learning voice signal and the standard deteriorated voice signal obtained for each short time period (hereinafter referred to as a frame) is equally divided into five divisions, and each division is divided into five divisions. The averaged similarity is input to the input layer of the neural network. The number of units in the intermediate layer is 5. The output layer with one unit has a learning voice signal MOS.
Enter Next, the back propagation method (D.
E. FIG. Rumelhart, J .; L. McClelllan
d, and the PDP Research Gro
up, Parallel Distributed P
processing Vol. 1 MIT Press,
pp. 318-362, 1986), the weighting coefficient between units in each layer is learned. A sigmoid function is used as the input / output function of each unit in the middle layer and the output layer. The quality of the learning sample and the non-learning sample is estimated for each learning process, and the difference between the quality value and the subjective quality measurement value is obtained. The processing ends when the difference between the quality estimation value of the learning sample and the subjective quality is stable at a small value, and the difference between the quality estimation value of the non-learning sample and the subjective quality becomes minimal, and the processing is ended. The weighting factors between units obtained from the series of learning processes are accumulated in the weighting factor database and used in the quality estimation process. The quality estimation processing equally divides the time series of the similarity of each frame of the test speech with respect to the standard deteriorated speech signal into five sections, and averages the similarity for each section with the same structure as the neural network used in the learning processing. It is input to the neural network and the weighting coefficient between the units obtained by the learning process and the sigmoid function are used to determine the MOS of the test voice.

[0004]

The speech quality objective estimation apparatus described above uses a speech signal of the same coding method as the test speech signal as a learning sample to determine the standard deteriorated speech and the weighting coefficient between units. Therefore, the quality can be estimated with sufficient accuracy. However, when estimating the quality of a test speech signal whose coding scheme is not clear, this speech quality estimation device may use a standard degraded speech signal and a weighting coefficient that are not compatible with the coding scheme. If a standard degraded speech signal that does not conform to the coding method is used, the degree of similarity between the standard degraded speech signal and the test speech signal cannot be obtained accurately. In addition, since the weighting coefficient learned by the neural network depends on the encoding method, a correct quality estimation value cannot be obtained when the encoding method is different.

According to the present invention, by incorporating a portion for identifying a voice signal coding system into this speech quality objective estimation device, a standard deteriorated voice signal and a weighting coefficient suitable for the test voice coding system are selected. It is an object of the present invention to provide a speech quality objective estimation device with good speech quality estimation accuracy that solves the above problems.

[0006]

[Problem to be Solved] A similarity between a standard deteriorated voice signal and a test voice signal is obtained, and a time series of the similarity is input to a neural network to objectively estimate the speech quality of the test voice signal. In the speech quality objective estimation apparatus, the speech analysis unit 2 for frequency-analyzing the test speech signal is provided, and the standard degraded speech database 5 for preliminarily storing various degraded speech signals is provided. A standard speech signal selected from the standard analysis speech database 5 and the analysis result of the speech analysis section 2 based on the identification result of the encoded voice identification section 1; The similarity calculation unit 3 for obtaining the time series of the similarity between the two is provided, and the result obtained by performing the learning processing by the neural network by the subjective evaluation value of various deteriorated speech signals and the similarity time series is obtained. And a weighting coefficient database 6 for accumulating weighting coefficients between sets in advance, and based on the time series of the similarity calculated by the similarity calculating section 3 and the weighting coefficient database 6 based on the discrimination result of the encoded speech discriminating section 1. The speech quality objective estimation device is configured to include the quality estimation value calculation unit 4 that constitutes the neural network that estimates the speech quality by inputting the selected weighting factor.

Then, the coded voice identification unit 1 frequency-analyzes the test voice signal, determines the spectrum outline of the test voice signal based on the analysis result, and identifies the coding system of the test voice signal by the spectrum outline. We constructed an objective speech quality estimation device.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION In an embodiment of the speech quality objective estimation apparatus of the present invention, first, a test speech is subjected to fast Fourier transform for each frame, and then the obtained spectrum is equally divided into several bands. , Average the spectra for each section. The averaged values are further averaged over the voice length, that is, the total number of frames. Then, the difference between the average of the lowest band spectrum and the full band spectrum is divided by the standard deviation of the full band spectrum. The value obtained by the division is defined as "a physical quantity used for identification", and the encoding method is identified by comparing this physical quantity with a threshold value. The threshold used for identification is determined by the distribution of the "physical quantity used for identification" obtained from the sample of the speech signal in which the code error has occurred.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of the present invention will be described with reference to FIG. In FIG. 1, reference numeral 1 denotes a coded voice identification unit added according to the present invention, which identifies a coding system of a coded voice signal in which a test voice signal is input and a code error occurs. The standard voice signal is selected from the standard voice database 5 using the discrimination result of the encoded voice discriminating unit 1, and the weighting factor obtained by the neural network is selected from the weighting factor database 6. The voice analysis unit 2 obtains the LPC cepstrum coefficient of the test voice signal for each frame. The similarity calculation unit 3 performs pattern matching between the test speech signal and the standard deteriorated speech signal selected from the standard deteriorated speech database 5 according to the identification result of the encoded speech identification unit 1 for each frame to determine the similarity. Find the time series. Here, the physical quantity representing the characteristics of the test voice signal and the standard deteriorated voice signal is the LPC cepstrum coefficient. To the quality estimation value calculation unit 4, the time series of the similarity is equally divided into 5 sections, the average of the degrees of similarity belonging to each section is input, and the encoding method identification is performed from the weight coefficient database 6. The weighting factor corresponding to the result is input, and the quality (MOS: average opinion score) is obtained. The standard deteriorated voice database 5 is a database of deteriorated voice signals representing various deteriorated voice signals. The weighting coefficient database 6 is a database of weighting coefficients between units obtained as a result of performing a learning process by a neural network using subjective evaluation values of various deteriorated speech signals and similarity time series.

FIG. 2 is a flow chart of the processing of the coded speech identifying section 1. First, in step 7, the test speech signal is fast Fourier transformed at 512 points for each frame to obtain a 256-point spectrum. In Step 8, since the target is telephone band voice (3400 Hz or less),
A 225-point spectrum of the obtained 256-point spectrum is equally divided into 9 bands and averaged for each band. Step 9
In the above, the entire 9-band spectrum is averaged over the voice length, that is, the total number of frames. In step 10,
The average difference between the spectrum of the lowest band and the spectrum of all 9 bands is divided by the standard deviation of the spectrum of all 9 bands. The value obtained by this division is taken as the "physical quantity used for identification", and the encoding method is identified by comparing this physical quantity with a threshold value. That is, in step 11, the “physical quantity used for identification” which is the value obtained in step 10 and the error rate of three levels (BER = 0, 10 ⁻³ , 1)
BER = 1 from the set of the ADPCM coded voice signal and the LDCELP coded voice signal in which a code error has occurred in 0 ^-2 ).
0 ^-2 ADPCM, and compares the threshold value for identifying the LDCELP encoded audio signal. As a result of comparison, "1" when the test speech signal is an ADPCM coded speech signal with BER = 10 ^-2 , and "^2" when the test speech signal is an LDCELP coded speech signal with BER = 10 ^-2 , Otherwise "3",
The identification result is output to the similarity calculation unit 3 and the quality estimation value calculation unit 4. In the similarity calculation unit 3 and the quality estimation value calculation unit 4, when the identification result is “1”, the similarity calculation unit 3 and the quality estimation value calculation unit 4 calculate the ADPCM coded speech with BER = 10 ^−2. Select standard degraded speech data and weighting factors that estimate the speech quality of the signal. Similarly, when the identification result is “2”, the standard deterioration voice data for estimating the speech quality of the LDCELP encoded voice signal of BER = 10 ⁻² from the similarity calculation unit 3 and the quality estimation value calculation unit 4 and Select a weighting factor. If the identification result is "3",
From the similarity calculation unit 3 and the quality estimation value calculation unit 4, BER =
10 ^-2 ADPCM coded audio signal and LDCELP
Standard degrading speech data and weighting factors for estimating speech quality other than the coded speech signal are selected.

Three-step error rate (BER = 0, 10 ^-3 , 1
ADPCM coded speech with a code error of 0 ^-2 ),
BER for LDCELP encoded voice 576 sentences
= 10 ⁻² ADPCM coded speech signal, LDCELP coded speech signal were identified, and BER = 1.
The discrimination correct answer rate of the LDCELP coded voice signal of 0 ^-2 is 9
7.9%, the identification correct answer rate of ADPCM coded voice signal of BER = 10 ^-2 is 83.3%, the identification correct answer rate of other voice signals is 75.0%, and the average of 3 categories is 87.3%. became. In addition, as a result of the quality measurement experiment in which the identification unit is incorporated, the accuracy is improved by about 0.25 as compared with the case where the identification unit is not incorporated.

[0012]

As described above, the speech quality objective estimation apparatus using the pattern matching and the neural network is provided with a portion for identifying the speech signal encoding method to identify the test speech encoding method. As a result, the call quality estimation is performed using the standard deteriorated voice signal and the weighting coefficient that are suitable for the test voice coding method, so that the accuracy of the call quality estimation can be improved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a call quality objective estimation device.

FIG. 2 is a diagram illustrating a procedure of processing of a coded voice identification unit.

[Explanation of symbols]

 1 coded speech identification unit 2 speech analysis unit 3 similarity calculation unit 4 quality estimation value calculation unit 5 standard deteriorated speech database 6 weighting coefficient database

Claims (2)

[Claims] 1. A call quality objective for objectively estimating the call quality of a test voice signal by obtaining a similarity between a standard degraded voice signal and a test voice signal and inputting a time series of the similarity into a neural network. The estimation device is equipped with a speech analysis unit for frequency-analyzing the test speech signal, equipped with a standard degraded speech database that stores various degraded speech signals in advance, and identifies the coding method of the coded speech signal in which a code error has occurred. A coded speech identification unit is provided, and a time series of the similarity between the analysis result in the speech analysis unit and the standard degraded speech signal selected based on the identification result of the encoded speech identification unit from the standard degraded speech database is obtained. Equipped with a similarity calculation unit, the weighting factors between units obtained as a result of learning processing by a neural network using subjective evaluation values of various deteriorated speech signals and similarity time series are stored in advance. The weighting coefficient database is provided, and the weighting coefficient selected from the time series of the similarity determined by the similarity calculating section and the weighting coefficient database based on the identification result of the encoded speech identification section is input-processed to improve the speech quality. An objective speech quality estimation apparatus comprising a quality estimation value calculation unit forming a neural network for estimation. 2. The speech quality objective estimation device according to claim 1, wherein the coded voice identification unit frequency-analyzes the test voice signal, and determines a spectrum outline of the test voice signal based on the analysis result. A speech quality objective estimation device characterized by identifying a coding system of a test voice signal by a rough shape.