JPS5917438B2 - Spectral parameter encoding method and speech synthesis device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a spectral parameter encoding method and a speech synthesis device in a speech analysis and synthesis system, in which speech is synthesized from parameters whose difference code has been reduced to zero.

The analysis and synthesis method uses a fixed-length window function,
For example, parameters expressing the spectrum information of the voice (spectral parameters) and parameters expressing the sound source information (sound source parameters) are separated and extracted from a finite number of data cut out by multiplying by a 30ms long window, etc. , the original audio signal is restored using the extracted parameters.

At this time, each parameter is extracted while moving the analysis window by a certain length of time (for example, 5 layers). The parameters extracted in this manner are encoded in one pass and transmitted or stored, read out from the receiving side or a storage device, decoded, and the decoded parameters are used to synthesize the original voice.

Parameters representing audio spectrum information are 8~
10 linear prediction coefficients or partial autocorrelation coefficients (PA
RCOR coefficient), etc. Linear prediction coefficients require around 10 bits of information per parameter when encoding, but PARCOR coefficients require around 10 bits of information depending on the order.
The amount of information required is ~4 bits, so there is an advantage that the original voice can be restored with less amount of information than using linear prediction coefficients. The present invention aims at compressing information on spectral parameters by differentially encoding spectral parameters and synthesizing speech from the differentially encoded parameters in such an analysis and synthesis method.

An example using PARCOR coefficients up to the 10th order will be described below. PARCOR coefficient Ki (1-1, 2, 3
, . . . 10) is a value in the range of 1, but in order to encode it, it is first necessary to perform quantization.

The number of bits required for quantization varies depending on the order of Ki; low-order PARCOR coefficients with high spectral sensitivity require a large number of bits, and high-order coefficients with low sensitivity require a small number of bits. For example, the spectral distortion per parameter should be 0.3 dB or less. The number of quantization bits required to make 10 {Ni}. 1=1 becomes as follows.

In the present invention, in order to reduce the number of quantization bits without deteriorating the quality of synthesized speech, the difference in spectral parameters is quantized and then encoded.

FIG. 1 shows the experimentally determined probability density function of the absolute value 1Δk1 of the change (difference) in Ki.

In Figure 1, 1 is the probability density function of lΔKll, 2 is iΔK
The probability density function of 2l to 1ΔKlOl is shown. ! Δ
K2l-1Jk10] distributions match extremely well and can be well approximated by the same exponential function. The distribution of IJkll concentrates in the range of 0 to 0.04, and can be approximated by an exponential function in this range, but there is also a non-negligible distribution in the range of 0.04 to 0.25. However, the first
It can be roughly approximated by combining two exponential functions as shown in 1 in the figure. From these probability density functions,
A nonlinear quantization characteristic that minimizes the quantization error of Ki can be determined.

For example, the 5-bit quantization characteristic conforming to 2 in Figure 1 is
It will look like the figure. By giving a quantization characteristic suitable for 2 in Fig. 1, the restoration accuracy of Ki is about the same as the bit allocation in equation (1) (
or less) such that the differential quantization bits {N
Di}. - If you look for it, it will look like the example 1:1.

This results in a reduction of 14 bits compared to the bit allocation in equation (1), and the parameter information is compressed by 23.7%.

Moreover, since the parameter restoration accuracy is on the same level (or less) as in the case of equation (1), the spectral distortion caused by the parameter quantization error can also be considered to be on the same level (or less). However, according to a simulation experiment, it was found that the synthesized speech sounds like 11 koro koro 1 locally, albeit slightly. From detailed data analysis, these 6? It was found that for the 11 sounds of Korokoro, the change in low-order PARCOR coefficients when transitioning from unvoiced to voiced sounds is large, resulting in poor restoration accuracy of spectral parameters.

Figure 3 shows the reconstruction error of k1, where the error is 0.1
It shows that it extends to

In the present invention, differential quantization is stopped and direct quantization is performed only when the spectral parameter restoration error exceeds a certain value.

FIG. 4 shows the block lines of the encoding processing device. In FIG. 4, 1 is a subtracter, 2 is a nonlinear quantizer, 3 is an adder,
4 is a delay memory, 5 is a linear quantizer, 6 is a comparator, 7 is a changeover switch, and 8 is an encoder. The spectral parameter Kif of the f-th frame (i is the parameter order, f is the frame number) and the parameter k1 of the previous frame (f-1
) (value including quantization error * difference), that is, ΔK
if is quantized by a non-linear quantizer 2 to obtain ΔKif, and the delay memory 4's ba, (no 1) is added to this by an adder 3 to obtain a restored value Kif of Kif.

The comparator tree 6 compares Kif and Kif and calculates the restoration error force '
When it is below a fixed value, ΔKlf is input to the encoder 8 by controlling the changeover switch 7, and when it exceeds the fixed value, it is input to the linear quantizer 5.
The linearly quantized Kif, that is, Kif, is sent to the encoder 8
Enter.

The encoder 8 uses *'ΔK with a predetermined code length for each parameter order.
If Kif is input, it is converted to a double-length code following the control code.

For example, as in equation (2), for ΔK2, ND2-5
When bits are assigned, the control code is 10000 (Δ
This code is not used when encoding K2. )
, K2 is represented by two codes of length following this code, that is, a 10-bit code. Furthermore, when the error exceeds a certain value, the input of the delay memory 4 tree' is also changed to Kif instead of Kif.

As a second means, difference quantization may be stopped only when the absolute value of the difference 1ΔKil exceeds a certain value, and direct quantization may be performed.

In this case, the only difference is the comparison target of the comparator 6 in FIG.
The other processing is exactly the same as the first means described above. The first means will be explained in more detail below. The procedure for restoring the original spectral parameters from the code Cif obtained by the first method will be described below.

First, for a code sequence {Cif} with code length NDl (bits), if each code Ci is not a control code, it is regarded as a differential code and ΔKif is restored, and the parameter value is calculated using the following equation. .

When a control code is detected, the following 2NDi (bits) are directly regarded as a code, Ki'f is restored, and this is used as a parameter value.

That is, the evaluation of the restoration error, which is the basis for switching from differential encoding to direct encoding, can be performed based on, for example, the quantization error resulting from the quantization bit allocation of equation (1). Since the possible value of Ki is -KiI minus 1, for example, for k1, 2/22 may be used as a comparison target for the restoration error of K. Next, a device for synthesizing speech from encoded parameters will be described.

FIG. 5 is a block diagram of a speech synthesizer according to the present invention. In FIG. 5, reference numeral 9 denotes a storage unit that stores sound source parameters and spectrum parameters extracted by analyzing the audio signal. The spectral parameters are encoded and stored using the above-mentioned means. A decoder 10 decodes the sound source parameters and outputs sound source control information, that is, pitch period P and amplitude A.

A decoder 11 decodes the encoded spectral parameter tree and outputs the difference ΔKif or the parameter value Kif.

Reference numeral 12 denotes an adder for performing the calculation of equation (3).

That is, the contents of the accumulator register 14 *
*Ki(f
-1) and the output ΔKif of the decoder 11 are
*Add and output parameter value Kif.

13 is a * switching gate, which normally sets the output Kif of the adder 12 to the accumulator register 14;
Only when a control code is detected at , the gate is switched and the output Krf from the decoder 11 is directly set in the accumulator register 14.

15 is a second switching gate, which transfers each parameter to the interpolation circuit 1.
6.

16 is an interpolation circuit for interpolating each parameter within a frame in order to synthesize smooth and natural speech.

17 is a white noise source for synthesizing unvoiced sounds, and PG is a periodic sound source for synthesizing voiced sounds.

The period of the voiced sound source is controlled by a pitch parameter Pf. Reference numeral 19 denotes a sound source switching gate, which switches the output of the white noise source 17 when the pitch parameter Pf is zero;
When it is 0, the output of the sound source 18 is input to the multiplier 20.

The multiplier 20 is for controlling the amplitude of the sound source using the amplitude parameter Af. 21 is a synthetic filter, which has been restored.

- *Controlled by spectral parameter Kif.

22 is a DA converter that converts synthesized speech into an analog quantity.

As described above, in the present invention, the spectral parameters are differentially quantized, and direct quantization is used only in areas with poor restoration accuracy, so that extremely high quality speech can be synthesized from a small amount of information.

Therefore, it is possible to realize an economical voice composing device.

[Brief explanation of drawings]

FIG. 1 is a characteristic diagram showing the probability density function of the difference value of spectral parameters, and FIG. 2 is a characteristic diagram showing the nonlinear quantization characteristic for optimally quantizing the difference in spectral parameters.
FIG. 3 is a graph showing the restoration accuracy of the spectral parameter k1, FIG. 4 is a block diagram of the spectral parameter encoding method of the present invention and the encoding processing device in the speech synthesizer, and FIG. 5 is a block diagram of the same speech synthesizer. FIG. 1...Subtractor, 2...Nonlinear quantizer, 3...One adder, 4...Delay memory,
6...Comparator, 5...Linear quantizer,
7... Selector switch, 8... Encoder,
9... Storage unit, 10, 11... Decoder, 12... Adder, 13, 15...
Switching gate, 16... Interpolation circuit, 17, 18...
...Sound source, 19...Sound source switching gate, 2
0... Multiplier, 21... Synthesis film, 22... DA converter.

Claims (1)

[Claims] 1 Spectral parameters extracted by analyzing speech are differentially encoded, and only when the restoration accuracy exceeds a value determined for each parameter order, differential encoding is stopped and parameter values are directly encoded. A spectral parameter encoding method characterized by: 2 When the difference value between two adjacent frames of spectral parameters extracted by analyzing the audio is less than a certain value determined for each parameter order, the difference value is encoded, and when the difference value exceeds a certain value, the parameter value is encoded. A spectral parameter encoding method characterized by direct encoding. 3 Analyze the audio and extract the extracted spectral parameters, and only when the restoration accuracy exceeds the value determined for each parameter order, stop the differential encoding and directly encode the parameter values, or analyze the audio and extract them. When the difference value between two adjacent frames of the spectral parameter obtained is less than a certain value determined for each parameter order, the difference value is encoded,
When the difference value exceeds a certain value, a memory for storing the spectral parameter obtained by directly encoding the parameter value, a decoder for reading and decoding the encoded parameter, and an addition for sequentially adding the decoded difference values. and an accumulator register, and a switching gate that sets the adder output to the accumulator register when the production value is decoded, and sets it directly to the accumulator register when the parameter value is decoded. Speech synthesis device.