JPH0211920B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a method for compressing parameter information in a speech analysis and synthesis method. What is the analysis and synthesis method? Parameters (spectral parameters) that express the spectrum information of the voice and sound source information are expressed from a finite number of data extracted by multiplying the discrete voice signal by a window function of a certain length, such as a 30ms long Hamming window. This method separates and extracts the parameters (sound source parameters) that are used in the audio signal, and then uses the extracted parameters to restore the original audio signal.
At this time, each parameter is extracted while moving the analysis window by a certain length of time. For example, the analysis window movement time length is typically 5ms, 10ms, or
A fixed value of 20ms is used. The parameters extracted in this way are encoded and transmitted or stored, read out from the receiving side or a storage device, decoded, and the original voice is synthesized using the decoded parameters. There are 8 to 10 linear prediction coefficients, partial autocorrelation coefficients (also called PARCOR coefficients), etc. as parameters representing the spectral information of speech.
Linear prediction coefficients require around 10 bits of information per parameter when encoding.
PARCOR coefficients require only 10 to 3 bits of information depending on the order, and therefore have the advantage that the original speech can be restored with less information than using linear prediction coefficients. The present invention provides a method for further compressing parameter information in such an analysis and synthesis method, and aims to realize an efficient analysis and synthesis method. This will be explained below along with an example using the PARCOR coefficient. The PARCOR coefficient k _i (i=1, 2, 3..., 10) is |
values in the range k _i <1, but quantization is required to encode them. The spectral distortion caused by quantization varies depending on the order of k _i , and the quantization error of K ₁ has the largest effect on the spectral distortion.
As the order increases, the influence decreases. The effect that fluctuations in K _i have on the spectrum is expressed as spectral sensitivity. The spectral sensitivity s ^(k) _ii is defined by the following equation. s ^(k) _ii = lim ΔKi→o | Δs / ΔK _i | ...(1) Here, Δs is the change in spectrum, and ΔK _i is
This is the change in the PARCOR coefficient K _i . PARCOR coefficients with high spectral sensitivity require many bits for quantization, while coefficients with low sensitivity require fewer bits. Proceedings of the Acoustical Society of Japan, 3-2-21, 1972
According to "Spectral Sensitivity Analysis of LPC Parameters" (Reference 1) published in May, when K _i is linearly quantized to N _i bits, the quantization distortion per parameter is σ _dB.
(dB scale display), it is shown that N _i =log ₂ √ ^(k) _ii −log ₂ (0.282σ _dB ) ...(2). _{From this , N i}
If (i=1, 2, . . . , 10) is determined, it can be considered that the influence of a change of 1 LSB on the spectral distortion is almost the same for any parameter. The present invention utilizes such properties of spectral parameters to reduce parameter information. Now, a value obtained by quantizing each parameter into N _i bits (in two's complement representation) is set as Q _i , and a parameter vector P is defined for each analysis frame by the following equation (3). P=(Q ₁ , Q ₂ , Q ₃ , . . . , Q _P ) (3) Furthermore, the distance D between vectors of two adjacent frames is defined by the following equation. D=| _PL −P _R |= _p 〓 ⁱ⁼¹ |Q _iL −Q _iR | (4) where p is the parameter order, L is the left frame, and R is the right frame. In the present invention, parameter information is compressed by using D as a measure and replacing parameter vector _PR with P _L only when D is less than a certain value. The figure is an explanatory diagram of parameter vector replacement operation. 1, 2, 3, . . . are frame numbers, and P _i (i=1, 2, 3, . . .) in the upper row is a spectral parameter vector extracted by analyzing the audio signal. The middle row shows an example of calculating D, and the binary number is
It is converted to decimal and displayed. The lower row shows the case where the parameter vector is replaced with the parameter vector of the previous frame when D is 20 or less. In the first frame, use P ₁ as is, and in the second frame
Since D is 4 in the frame, P ₂ is replaced with P ₁ . In the third frame, the distance D from the vector P ₁ of the second frame after replacement is 12, so
Replace P ₃ with P ₁ . If we calculate D for the fourth frame in the same way, it will be 20 or more, so
Use P ₄ as is. In the fifth frame, D
Since it becomes 10, replace P ₅ with P ₄ . In the 6th frame, D becomes 26 and exceeds 20, so P ₆ is used as is. In this way, when the replacement with the vector sequence shown in the lower part of the figure is completed, the parameter information for 6 frames can be expressed by the parameter information for 3 frames, and the amount of information can be greatly reduced. Journal of the Institute of Electronics and Communication Engineers, 78/2, Vol.61-A, No.
2. According to ``Optimal Code Configuration in PARCOR-type Voice Analysis and Synthesis System'' (Reference 2), when only one of the spectral parameters k _i is quantized to N _i bits, the quantization distortion DS ^(p) _Qi It has been shown that the empirical formula (5) holds between (N _i ) and the total quantization distortion DS ^(p) _Q (N) when each of k _i is quantized to N _i bits. DS ^(p) _Q (N)= _p 〓 ⁱ⁼¹ DS ^(p) _Qi (N _i ), (dB) ² ...(5) Here, N= _p 〓 ⁱ⁼¹ N _i , (bit) ... ...(6), p is the parameter order, and in the example p=
It is set at 10. In the present invention, DS ^(p) _Qi in equation (5)
Since (N _i ) is set to a constant value (σ _dB ) ² regardless of the value of i, DS ^(p) _Q (N) = _p 〓 ⁱ⁼¹ (σ _dB ) ² = p(σ _dB ) ² ...(7) becomes. On the other hand, if N _i determined by equation (2) is used, the effects of changes in Q _i (i = 1, 2, ..., 10) in equation (3) on spectral distortion can be considered to be equivalent, so Adding the distortion [D=(σ _dB ) ² ] that occurs when the parameter spectrum _PR is replaced by _PL to the quantization distortion to find the total amount of spectral distortion √, we get √=√()+( _dB ) ² =√(+)・σ _dB , (dB)...(8) Distance between vectors D defined by formula (4)
It can be seen that is a measure that indirectly expresses spectral distortion using equation (8). That is, if the value of D is selected to a certain value, P _R
By replacing PL with _PL , the spectral distortion caused can be kept below a certain value. According to a simulation experiment conducted with an analysis window length of 30 ms, a frame period of 10 ms, and p = 10, σ _dB
{N _i } ¹⁰ _i=1 calculated as = 1 dB is {N _i } ¹⁰ _i=1 = {7, 6, 5, 4, 4
, 4, 3, 3, 3, 3}...(9), and the results of replacing the parameter vector P with D=0, 10, 20, 30 are as shown in Table 1. Experiments have shown that when a replacement operation is performed on an unvoiced frame, the phonology deteriorates significantly, but good results are obtained for consecutive voiced frames. The following table summarizes the latter case.

[Table] When D=30, the deterioration in quality is noticeable audibly, but up to D=20, it is almost as good as when D=0. It can be seen that when D=20, information is reduced by 33% compared to D=0, and extremely efficient analysis and synthesis can be achieved. As is clear from the above description, in the present invention,
Before calculating the distance between parameter vectors, each spectral parameter is quantized so that the spectral distortion per spectral parameter is equal, so the calculation result of the distance between each parameter spectrum has the same weight. Spectral parameters are replaced with values from one frame before, as long as the influence on spectral distortion is equal and spectral distortion does not exceed a certain value using the distance between vectors that indirectly expresses this spectral distortion as an evaluation measure. Since the amount of information is reduced by using the method, it has the effect of realizing extremely efficient analysis and synthesis. Furthermore, if parameters compressed in this way are used in a speech synthesis device that stores them in a storage device, the memory cost can be greatly reduced and an extremely economical device can be realized.

[Brief explanation of drawings]

The drawing is a diagram illustrating a parameter vector replacement operation using the parameter information compression method according to the present invention.

Claims (1)

[Claims] 1 After quantizing each spectral parameter so that the spectral distortion per audio spectral parameter is equal, find the distance between the parameter spectra of two adjacent frames, and when the distance between the parameter vectors is less than a certain value, A parameter information compression method characterized by replacing parameters of a subsequent frame with parameters of a previous frame.