JPH06195093A - Automatic tone quality evaluating device - Google Patents

Abstract

PURPOSE:To find a predictive subjective evaluated value which has high correlation with a subjective evaluated value even for an encoded speech having <=8Kb/s low bit rate when tone quality is evaluated by using feature parameters extracted from a speech signal. CONSTITUTION:This device is equipped with an encoding and decoding part 3 which generates a reproduced speech signal by encoding and decoding an input speech signal, an input speech signal feature extraction part 4 which extracts feature parameters from the input speech signal, a reproduced speech signal feature extraction part 5 which extracts feature parameters from the reproduced speech signal, an input speech signal dynamic feature extraction part 6 which extracts dynamic feature parameters by using the feature parameters of the input speech signal, a reproduced speech signal dynamic feature extraction part 7 which extracts dynamic feature parameters by using the feature parameters of the reproduced speech signal, an objective evaluation part 8 which outputs the distances between the dynamic feature parameters of the input speech signal and the dynamic feature parameters of the reproduced speech signal, and a subjective evaluation and prediction part 9 which predicts the subject evaluated value by using an objective evaluated value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound quality evaluation device,
In particular, the present invention relates to an apparatus for predicting a subjective evaluation value of a reproduced voice signal from the distance between the characteristic parameter of the input voice signal and the characteristic parameter of the reproduced voice signal.

[0002]

2. Description of the Related Art Conventionally, in order to evaluate the sound quality of a voice, in addition to subjectively evaluating the evaluation voice in a listening experiment, it is extracted from the original feature voice of the voice and the evaluation voice to obtain the ratio and distance between the feature parameters. There is a method of objective evaluation. For objective evaluation, in addition to the SN ratio (Japanese Patent Laid-Open No. 63-273895) (Reference 1) and the segmental SN ratio, the cepstrum distance and BS
D distance (Proc.ICASSP 91, IEEE S
Peach Processing. , Vol. 1, p
p. 493-496, 1991) (reference 2) and the like, spectral distortion is used. Recently, a model for predicting a subjective evaluation value from the characteristic parameters of the voice used for the objective evaluation has also been studied (Journal of the Institute of Electronics, Information and Communication Engineers A Vol. J73-A No. 6 p.
p. 1039-1047 June 1990) (Reference 3).

[0003]

However, these conventional sound quality evaluation methods have a problem that it is difficult to correspond to a subjective evaluation value for coded speech having a low bit rate of 8 kb / s or less.

[0004]

In order to solve the above problems, an automatic sound quality evaluation apparatus according to the first invention inputs a voice signal and encodes / decodes the input voice signal to generate a reproduced voice signal. An encoding / decoding unit for extracting the characteristic parameter from the input audio signal, a reproduced audio signal characteristic extracting unit for extracting the characteristic parameter from the reproduced audio signal, and an input audio signal characteristic extracting unit for extracting the characteristic parameter from the reproduced audio signal. An input voice signal dynamic feature extraction unit that extracts a dynamic feature parameter using a feature parameter; a playback voice signal dynamic feature extraction unit that extracts a dynamic feature parameter using a feature parameter of the playback voice signal; A first objective evaluation unit for outputting a predetermined distance between a dynamic characteristic parameter of an input audio signal and a dynamic characteristic parameter of the reproduced audio signal; and the first objective evaluation. Using an objective evaluation value obtained by having a subjective evaluation prediction section for predicting a subjective evaluation value of the reproduced audio signal.

Alternatively, the automatic sound quality evaluation apparatus according to the second invention is the automatic sound quality evaluation apparatus according to the first invention, further comprising a second objective evaluation section for outputting a distance between the characteristic parameter of the input audio signal and the characteristic parameter of the reproduced audio signal. A composite subjective evaluation prediction unit that predicts a subjective evaluation value by using the objective evaluation value obtained by the first objective evaluation unit and the objective evaluation value obtained by the second objective evaluation unit .

Alternatively, the automatic sound quality evaluation apparatus according to the third invention is the automatic sound quality evaluation apparatus according to the first invention, wherein the reproduced sound coded / decoded by a predetermined coding method is input instead of the coding / decoding unit. It has a reproduction voice input terminal for.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to FIGS. FIG. 1 is a block diagram showing an embodiment of the sound quality evaluation apparatus of the first invention.

In FIG. 1, a voice signal is input from an input terminal 1 and sent to an encoding / decoding unit 3 and an input voice signal feature extraction unit 4. The encoding / decoding unit 3 performs encoding / decoding using the original signal to create a reproduced voice signal. For encoding / decoding, for example, CELP (Proc.
nt. Conf. Accout. , Speech, Si
general Processing, pp. 200-20
3, 1989) (reference 4) and the like. The reproduced audio signal generated by the encoding / decoding unit 3 is sent to the reproduced audio signal feature extracting unit 5.

The input voice signal feature extraction unit 4 uses the input voice signal S _x to obtain the feature parameter xpar at regular time intervals (frames). The feature parameter xpar obtained by the input voice signal feature extraction unit 4 is sent to the input voice signal dynamic feature extraction unit 6. As the feature parameter xpar obtained by the input voice signal feature extraction unit 6, well-known ones such as rms, Bark spectrum (Reference 2), pitch, and cepstrum can be used.

Here, as an example of obtaining the characteristic parameter xpar from the input voice signal S _x , rms and Ba will be described below.
The formula for obtaining the rk spectrum is shown.

First, an equation for obtaining the characteristic parameter xrms ^(k) of the k-th frame from the input voice signal will be shown.

[0012]

[Equation 1]

Next, the procedure for extracting the Bark spectrum from the input voice signal will be described.

1. FF for the input voice signal S _x ^(k)
T is performed to obtain the power spectrum X ^{( k)} (f).

2. Power spectrum X ^(k) (f) is B
Convert to ark scale Y ^(k) (x). The conversion from x to y is performed using the following relational expression.

F = 600 sinh (i / 6) 3. Power spectrum Y converted to Bark scale
^The critical band filter F ^(k) (x) is applied to ^(k) (x) to obtain the excitation patternD.
^(k) (x) is calculated. Critical filter F ^(k) (x)
Is represented by the following formula.

10 log ₁₀ F ^(k) (x) = 7-7.5 (x-α)
−17.5 [0.196+ (x−α) ² ] ^1/2 Here, α = 0.215. D ^(k) (x) = F ^(k) (x) * Y ^(k) (x) (* is a convolution operator.) D ^(k) (x) -Excitation pattern
3. N F ^(k) (x) -critical band filter Y ^(k) (x) -Bark scale converted power spectrum 4. The perceptual weighting is applied to the Excitation pattern D ^(k) .

Hearing weight H from 1800 to 3400 Hz
(F) can be obtained from the following equation.

H (f) = (2.6 + e -2 ^{j π f} ) /
(1.6 + e −2 ^{j π f} ) H (f) = 1 at 1800 Hz or lower, and the same value as 3400 Hz at 3400 Hz or higher.

[0020]

[Equation 2]

5. Bark spectrum {xB ^(k) [i], i = 1 ,. ． ． , I} is calculated by the following equation.

[0022]

[Equation 3]

However, channel number i and Bark
There is a relationship of x = 1.0 × i with Scale x.

The input signals rms and Ba obtained by the above procedure
The feature parameters xrms and xB of the rk spectrum are output to the input voice signal dynamic feature extraction unit 6.

The reproduced voice signal feature extraction unit 5 uses the reproduced voice signal S _y to obtain the characteristic parameter ypar for each fixed time (frame). The characteristic parameter ypar obtained by the reproduced voice signal characteristic extraction unit 5 is sent to the reproduced voice signal dynamic characteristic extraction unit 7. The characteristic parameters ypar obtained by the reproduced voice signal characteristic extraction unit 5 include, for example, characteristic parameter rms Bark spectrum (reference 2), pitch, cepstrum, and the like. The method for obtaining the characteristic parameter ypar of the reproduced audio signal from the reproduced audio signal S _y is as follows.
Since the method is the same as the method of obtaining the characteristic parameter of the input audio signal using the input audio signal S _x , the description thereof is omitted here.

The input voice signal dynamic feature extraction unit 6 will be described. Next, an operation of converting the characteristic parameter xpar of the input speech signal S _x obtained by the input speech signal characteristic extraction unit 4 into a dynamic characteristic parameter δxpar and sending it to the first objective evaluation unit 8 is performed. The audio signal dynamic feature extraction unit 6 will be described. Input voice signal dynamic feature extraction unit 6
Since there are several methods for converting the characteristic parameter xpar of the input speech signal into the dynamic characteristic parameter δxpar, an example will be given here.

[0027] First, the characteristic parameters XPAR of the input speech signal, the method for converting a dynamic characteristic parameter Derutaxpar, wherein parameters XPAR ₁ of the input audio signal,
Dynamic feature parameter δxpa converted from xpar ₁
An explanation will be given using r ₁ .

Dynamic feature parameter δxpar _{1 of} the feature parameter xpar ₁ ^(k) extracted in the k-th frame
^(k) is the feature parameter xpar ₁ before the 1-frame
Feature parameter xpar after 1 frame from ^{(k + l)}
Calculated by the difference of ₁ ^(k-l) .

Δxpar ₁ ^(k) = xpar _{1
^{(K + l) -xpar 1 (}} k - l) k - frame number l - frame shift number xpar ₁ ^(k) - characteristic parameters Derutaxpar ₁ of the k frame of the input audio signal ^(k) - k-th input speech signal Dynamic Feature Parameter of Frame A method of obtaining the dynamic feature parameter when the feature parameter xpar ₁ of the input speech signal is rms is shown below.
Here, the characteristic parameter of rms of the input voice signal is xr
m and the dynamic feature parameter is δxrms, the dynamic feature parameter δxrms ^(k) of the rms of the k-th frame
Can be calculated by the following formula.

Δxrms ^(k) = xrms ^{(k + l)} -
xrms ^(k-l) k-frame number l-frame shift number xrms ^(k) -rm of kth frame of input speech signal
s δxrms ^(k) -rm of the kth frame of the input speech signal
dynamic characteristic parameter of s then the characteristic parameters XPAR of the input speech signal, the method for converting a dynamic characteristic parameter Derutaxpar, when the characteristic parameter XPAR ₂ of the input audio signal is a multi-dimensional, is converted from XPAR ₂ Dynamic feature parameter δx
This will be described as par ₂ .

Characteristic parameter {xpar ₂ ^(k) [t], t = of the T-th dimension extracted in the k-th frame
1 ,. ． ． , T} dynamic feature parameter δ par ₂
^(k) [t] can be calculated by the following equation.

Δxpar ₂ ^(k) [t] = xpar ₂
^{(K + l) [t]} -xpar 2 (k - l) [t] t = 1 ,. ． ． , T k-frame number t-dimension number T-dimension number l-frame shift number xpar ₂ ^(k) [t] -kth frame of the input speech signal, characteristic parameter of the tth dimension δxpar ₂ ^(k) [t ] -Dynamic feature parameter of k-th frame and t-th dimension of input voice signal The method of obtaining the dynamic feature parameter when the feature parameter xpar ₂ of the input voice signal is the Bark spectrum is shown below. Here, the characteristic parameter of the Bark spectrum of the input speech signal is xB, and the dynamic characteristic parameter is δx.
Let B be the I-th dimension Bar extracted in the kth frame
Characteristic parameter of k spectrum {xB ^(K) [i], i
= 1 ,. ． ． , I} dynamic feature parameter δxB ^(K)
[I] can be obtained by the following formula.

ΔxB ^(k) = xB ^{(k + l)} [i] -x
B ^(k-1) [i] k-frame number i-dimension number l-frame shift number xB ^(k) -bark of k-th frame of input speech signal
Spectrum δxB ^(k) -Bark of k-th frame of input speech signal
Dynamic Feature Parameter of Spectral A method other than the above for converting the feature parameter xpar of the input speech signal into the dynamic feature parameter δxpar is used as the feature parameters xpar ₃ and xp of the input speech signal.
ar ₃ dynamic characteristic parameters transformed from Derutaxpar ₃
Use to explain.

Dynamic feature parameter δxp of the k-th frame
ar ₃ ^(k) is an average feature parameter avg− of the feature parameter xpar ₃ ^(k) and the feature parameter xpar _3.
It is obtained by the difference of xpar.

[0035]

[Equation 4]

Further, the characteristic parameter x of the input voice signal
As a method of converting the par into the dynamic feature parameter δxpar, the feature parameter xpar of the input audio signal is used.
₄ and the dynamic feature parameter δ converted from xpar ₄
Explain using xpar ₄ .

[0037]

[Equation 5]

The characteristic parameter ypar of the reproduced voice signal S _y obtained by the reproduced voice signal characteristic extraction unit 5 is
The reproduced voice signal dynamic feature extraction unit 7 converts the dynamic feature parameter δypar and sends the dynamic feature parameter δypar to the first objective evaluation unit 8. The method for obtaining the dynamic feature parameter in the reproduced voice signal dynamic feature extraction unit 7 is the same as that in the input voice signal feature extraction unit 6, and therefore the description is omitted here.

The dynamic feature parameters obtained by the input voice signal dynamic feature extraction unit 6 and the reproduced voice signal dynamic feature extraction unit 7 are supplied to the first objective evaluation unit 8 as dynamic feature parameters of the input signal. xpar and the dynamic characteristic parameter yp of the reproduced signal
The distance from ar is calculated and sent as an objective evaluation value δAVG of the dynamic feature parameter to the subjective evaluation prediction unit 9. The method of obtaining the objective evaluation value from the dynamic feature parameters is shown below.

First, a method of obtaining an objective evaluation value when the dynamic feature parameter is one-dimensional will be described. Here, assuming that the dynamic characteristic parameter of the input speech signal is δxpar ₁ , the dynamic characteristic parameter of the reproduced speech signal is δypar ₁ , and the objective evaluation value is δAVG ₁ , the dynamic characteristic parameter δxpar ₁ of the input speech signal of the k-th frame. ^(k) and the dynamic feature parameter δypar of the reproduced audio signal of the k-th frame
_The objective evaluation value δAVG ₁ with ₁ ^(k) is calculated by the following equation.

[0041]

[Equation 6]

Next, the dynamic feature parameter δx of the input voice
The par ₁ δxrms, δyrms dynamic characteristic parameter Derutaypar ₁ playback sound, when an objective evaluation value DerutaAVG ₁ of the dynamic characteristic parameter and DerutaAVG _rms, dynamic characteristic parameter of the rms of the input speech signal of the k frame δx
rms ₁ ^(k) and rm of the reproduced audio signal of the k-th frame
The objective evaluation value δAVG _rms with the dynamic feature parameter δyrms ₁ ^{(k) of s} is obtained by the following formula.

[0043]

[Equation 7]

Further, the method of obtaining the objective evaluation value when the dynamic feature parameter is multidimensional is as follows: the dynamic feature parameter of the input voice signal is δxpar ₂ , the dynamic feature parameter of the reproduced voice signal is δypar ₂ , and the objective assessment is performed. The value will be described as δAVG ₂ .

[0045]

[Equation 8]

Here, the dynamic feature parameter δ of the input voice
xpar ₂ is δxB, dynamic characteristic parameter δ of reproduced voice
A method of obtaining _BSD will be described, where ypar ₂ is δyB and the objective evaluation value δAVG ₂ of the dynamic feature parameter is δAVG _BSD . Dynamic characteristic parameter δxB the BSD of the input audio signal of the k-th frame ^(k), the dynamic characteristic parameter δyB the BSD of reproduced audio signal of the k-th frame ^(k), an objective evaluation value DerutaAVG _BSD following formula Required by.

The objective evaluation value δAVG _BSD can be calculated by the following formula.

[0048]

[Equation 9]

The objective evaluation value of the dynamic feature parameter obtained by the first objective evaluation unit 8 by the above method is sent to the subjective evaluation prediction unit 8.

The subjective evaluation prediction unit 8 predicts the subjective evaluation value with at least one objective evaluation value and at least two prediction coefficients, and outputs the evaluation result from the output terminal 2. The prediction coefficient is obtained so that the error between the subjective evaluation value and the prediction evaluation value collected in advance using a large amount of voice data is minimized. The relationship between the prediction coefficient a and the subjective evaluation value is shown below.

[0051]

[Equation 10]

The formula for obtaining the predicted subjective evaluation value using the prediction coefficient and the objective evaluation value obtained by the first objective evaluation unit 8 is shown below. The predicted subjective evaluation value is

[0053]

[Equation 11]

The prediction coefficients are a, b, and c, and the objective evaluation value of the dynamic feature parameter obtained by the first objective evaluation unit 8 is δA.
Let VGa and δAVGb.

[0055]

[Equation 12]

FIG. 2 is a block diagram showing an embodiment of the sound quality evaluation apparatus of the second invention. In FIG. 2, components having the same numbers operate in the same manner as in FIG.

In the second objective evaluation unit 10, the characteristic parameter xpar obtained by the input voice signal feature extraction unit 4 from the input voice signal S _x and the reproduced voice signal feature extraction unit 5 from the reproduced voice signal S _y are obtained. , With the feature parameter ypar,
The distance is calculated as the objective evaluation value AVG and output to the composite subjective evaluation prediction unit 12.

Here, the characteristic parameter of the input audio signal is xpar, the characteristic parameter of the reproduced audio signal is ypar,
When AVG objective evaluation value, the characteristic parameters xpar of the input speech signal of the k-th frame ^(k), the characteristic parameters ypar the reproduced audio signal of the k-th frame ^(k),
The objective evaluation value AVG is calculated by the following formula.

[0059]

[Equation 13]

In the composite subjective evaluation prediction unit 12, at least one objective evaluation value obtained by the first objective evaluation unit 8 and at least one objective evaluation value obtained by the second objective evaluation unit 10.
The subjective evaluation value is predicted with one objective evaluation value and at least three prediction coefficients, and the evaluation result is output from the output terminal 2. The operation of the composite subjective evaluation prediction unit 12 is the same as that of the subjective evaluation prediction unit 9, and thus the description thereof will be omitted.

FIG. 3 is a block diagram showing an embodiment of the sound quality evaluation apparatus of the third invention. In FIG. 3, components having the same numbers operate in the same way as in FIGS.

From the reproduced voice input terminal 6, the reproduced voice obtained by encoding / decoding with the same voice signal as the voice signal input from the input terminal 1 is input and output to the reproduced voice signal feature extraction unit 5. To do.

[0063]

As described above, the sound quality evaluation method according to the present invention converts a conventionally used feature parameter into a dynamic feature parameter, so that even a low bit rate voice is correlated with a subjective evaluation value. It is possible to perform high-quality sound evaluation.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a sound quality evaluation apparatus of the first invention.

FIG. 2 is a block diagram showing an embodiment of a sound quality evaluation apparatus of the second invention.

FIG. 3 is a block diagram showing an embodiment of a sound quality evaluation apparatus of the third invention.

[Explanation of symbols]

 1 Input Terminal 2 Output Terminal 3 Encoding / Decoding Section 4 Input Voice Signal Feature Extraction Section 5 Playback Voice Signal Feature Extraction Section 6 Input Voice Signal Dynamic Feature Extraction Section 7 Playback Voice Signal Dynamic Feature Extraction Section 8 First Objective Evaluation unit 9 Subjective evaluation prediction unit 10 Second objective evaluation unit 12 Composite subjective evaluation prediction unit 13 Playback audio input terminal

Claims (3)

[Claims] 1. An encoding / decoding unit for inputting an audio signal and encoding / decoding the input audio signal to create a reproduced audio signal; and an input audio signal feature extraction for extracting a characteristic parameter from the input audio signal. A playback voice signal feature extraction unit that extracts a feature parameter from the playback voice signal, an input voice signal dynamic feature extraction unit that extracts a dynamic feature parameter using the feature parameter of the input voice signal, and the playback A playback voice signal dynamic feature extraction unit that extracts a dynamic feature parameter using a feature parameter of a voice signal, and a predetermined dynamic feature parameter of the input voice signal and a dynamic feature parameter of the playback voice signal. A first objective evaluation unit that outputs a distance, and a subjective evaluation that predicts the subjective evaluation value of the reproduced audio signal using the objective evaluation value obtained by the first objective evaluation unit. A sound quality evaluation apparatus having a prediction unit. 2. The sound quality evaluation apparatus according to claim 1, wherein
Furthermore, a second objective evaluation unit that outputs a distance between the characteristic parameter of the input audio signal and the characteristic parameter of the reproduced audio signal, the objective evaluation value obtained by the first objective evaluation unit, and A sound quality evaluation apparatus comprising a composite subjective evaluation prediction unit that predicts a subjective evaluation value using the objective evaluation value obtained by the second objective evaluation unit. 3. The sound quality evaluation apparatus according to claim 1,
A sound quality evaluation apparatus comprising, instead of the encoding / decoding unit, a reproduced sound input terminal for inputting reproduced sound encoded / decoded by a predetermined encoding method.