JPS5979299A - Coding of voice analysis - 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 APPLICATION OF F'1 RAKIJU This invention relates to a method for converting audio data provided to a layer synthesis device d into an analysis tube.

Conventional configurations and their problems Until now, various methods have been used as voice analysis and synthesis methods, such as Ij+
! Among the two methods, the analysis and synthesis method using linear prediction coefficients or equivalent partial autocorrelation coefficients is inferior in terms of information compression, quality of synthesized speech, etc.

After pulling (1), the audio digital data (1) is turned on. Audio digital data (2) can be converted into linear reins or f! ]＋! Using autocorrelation and the like, analysis (3) is carried out for each analysis time width (hereinafter referred to as a frame) to determine the characteristics of the voice and obtain parameters (4). Here, the parameters (4) are the basic homogeneous wave number, amplitude, linear prediction coefficient or partial autocorrelation coefficient, voiced/unvoiced judgment value, total 1 (L power (RMS of human voice digital data, hereinafter referred to as E), residual power (rubfs of edge shape prediction error, hereinafter 7"
R and hail). The obtained parameter (4) is encoded (5) and the voice is synthesized (7) using the encoded parameter ([).

FIG. 2 shows a detailed explanation of the speech synthesizer used for synthesis (7) in FIG. 1. The audio parameters are stored in encoded form in the parameter ROM0:I, and the starting address of each word is stored in the address table (6). By inputting the number of voices you want to utter from the address designation end 0υ, select the parameter ROMfJ:9 in which the corresponding voice parameters are stored.
The address of is entered in the address table Q, and the address is
The parameter is 1f71i from meter ROM0!
1 out, 31■ro. The read parameters are bit-packed, so each 'n' F, 4s
The bits 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, , , , , , , , , , , , , , , , , , , , , , ,. Since 1X is written in
j: L) between the parameter and the parameter value one frame ago! , 1.1, and the actual analysis time width, 1. ri short time 11! Update the parameters with I lj'ilj, and perform the following operations, iil chibi'tsuchi/'J unta dry, voiced oil circuit (c), voiceless ff r!bj times i1
1? The pitch counter αη is set at 11·jl rill to the fundamental frequency when synthesizing the I/S voice, and is calculated every 1 period, that is, the fundamental 9 frequencies at each embedding frequency. 1] Add 1 (1) to the pulse that is the voice sound leech. No j: 1 sound source circuit 0 [
! generates random noise that becomes f'i'l'b:i when synthesizing unvoiced sounds. Digital filter (E)
is determined by a partial autocorrelation coefficient (hereinafter referred to as /f parameter and 1 value) and models the shape of the vocal tract. These are used to synthesize speech, and the D/A converter C1
) is converted to an analog signal, and a composite waveform is obtained at the output terminal (support).

The 30th is the 2nd ('' decoding ROM (1!1 of 4
This shows 71 formations. i J: Depending on how bits are given to each noterameter during encoding, 1;:;
different. In Figure 3, 10th order parameters are used, amplitude is 5 bits, fundamental frequency is 6 bits, K1-7 bits, K2-
G-Pill-1K3-5 bits, K4-K6,4.
bit, (7.about.1) shows the case of 10-18 bits.

FIG. 4 shows the configuration of parameters for one frame when the above bit allocation is given. For the fundamental frequency and l (parameters), if the value of a certain frame is the same as or close to the previous frame, the value of the previous frame is not memorized and the value of the previous frame is used as is.The repeat bit is set to 1. The bit star required for the parameter can be compressed.

Conventionally, fundamental frequency, amplitude, and I (parameters are encoded and decoded separately with different hit distributions for each order, but the spectral sensitivity of the K parameter is s i,,':'i
-n, there is only one sign for concentric parameters,
A reconstruction table is used.

In general, the residual '1if' and the force R are expressed as the sum of squares of the difference between the audio data predicted using the extracted parameters and the original waveform, indicating the accuracy of the extracted parameters.
The elementary schools are well known. That is, when the residual power R is large, the precision of the parameters of that frame is low and the distortion is large. Conversely, when the residual power R is small, it means that the parameters are extracted with high accuracy.

However, since the residual zrE j'lt force depends on the amplitude of the human power data, it is necessary to normalize it with the total power E of the input data. That is, when R/Iε is large, the accuracy of the parameters of the frame is low, and when R/E is small, the parameters of the frame have a high degree of M. For parameters with high precision, it is necessary to code υ with high bits in detail, but for parameters with low precision, even if the bit allocation is reduced, there is little effect on the synthesis quality.

Purpose of the Invention The purpose of the present invention is to take the above points into consideration and to propose a voice analysis encoding method that can improve the sound quality without increasing the data 1,1. It is something.

176 Achievement of the Invention In order to achieve the above object, the present invention determines +1 value 1), and when 1 and L<> H, low bit allocation, and I<7
'When E and H, perform encoding with high bit allocation, 1;
As a result, at the same bit rate as before,
It is possible to obtain a synthesized sound with higher crystalline y:j than before by giving t depth/shin bits to the less accurate parameter and encoding it, and when R/E<H, it is the same as before. Encoding is performed with bit allocation, and when R/E: profit is I, encoding is performed with less bit allocation than before, and the conventional crystal 'i'i;'r= can be obtained at a lower bit rate than before. have an effect.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below based on the drawings.
1 (Analyze the audio data sambling with
O.; When performing a 10th order partial autocorrelation analysis at 1 price,
Fig. 5(a)
As shown in (b), when VE > 0.4 and R/
Encoding is performed using different bit allocations for E-20 and E-4. For example, when R/)E < 0.4, the accuracy of the parameter is t'! Since the bit allocation is small, the conventional bit allocation is used: 5 bits for nobkQ'11i, 6 bits for the fundamental frequency, and 6 bits for the fundamental frequency.
1-7 bits, K 2-6 bits, K8-5 bits,
K 4, K6-4 bits, K 7-I (10' 8
When 1</s;> 0.4, the encoding is performed by reducing each parameter by 1 bit compared to the conventional method. That is, K1-6 bits, K25 bits,
J (8-4 bits, K4~6-3 bits, K7~KI
It is O-2 bit. The same bit allocation is used for amplitude and fundamental frequency. That frame is I</E>
0.4 or R/'Ii, <-0, .mu., ie, one bit is required as information for dividing the decoding table. This is called an encoding flag. Table 1 shows the percentage ratio of frames with R>0.4 and frames with R/'E<0.4 when analyzing a 20-second weather forecast. Based on this, we calculate the number of pins required to record one second of data. Conventionally, it takes 55 pins to store one frame, and since the analysis time span is lQm, it takes 1 sorting. In this case, it is necessary to store 100 frames, so 55 x 100 = 5500 bits/second, but in this method, + 47 x 88 + 56 x 67 = 58
08 bits/sho. In other words, compression of about 200 bits is possible with 4C, and the same level of synthetic output as before can be obtained.・Table 1, Figure 6 (a) and (b) show the configuration of parameters for one frame. Figure 6(a) shows the case of 1</E (0,4, 56 bits for one frame, Figure 6(b) shows 1v
1<>0.4, and one frame has 47 bits.

Note that here, we have omitted an example in which there is one ``1'' and two decoding tables are used; however, it may be divided into the same stages depending on the ratio of the ``1'' parameter N'+';ICj:.

Next i/(F7!
．． Regarding frame No. 4, since the parameter accuracy tC is low, the parameters of the previous frame are used as they are. At that time, the number of bits required for the frame is 1
2 bits is sufficient, and its jh'J configuration is shown in Figure 21)7.

Decoding (decoding) M has the same structure as the conventional one, as shown in FIG. In this example, the number of bits per second is 56 x 67 + 12 x 83 = 4
There are 648 bits/second i, and data compression of approximately 1850 bits/second is possible.

Effects of the Invention According to the Miki invention, since encoding is performed at a low bit rate only in areas where parameter extraction precision is low, data compression can be performed with almost no deterioration in sound quality. Furthermore, by allocating the reduced number of bits in areas with low extraction accuracy to areas with high extraction accuracy, the data size is increased.

You can easily improve the sound quality.

[Brief explanation of the drawing]

Figure 1 is a flowchart of the speech analysis and synthesis method -/', Figure 2 is a block diagram of the speech synthesizer, and Figure 81-1l is a '4r (E
Next CIIG'l-decoding 1 <OM's ri
An example of I'j composition C71, Ff'41 is an example diagram of (lt composition) of parameters for one frame in the conventional example.
The figure is part 7Iil of the present invention! Decoding R (JM configuration 1, 5" f5.6 figure is rl'i
5 +21 (parameter configuration for 1 frame in 2 [i, 1'57 Figure 1.1 This is a configuration diagram of 1 frame's worth of parameters in another embodiment. θ...Parameter ROM , (+(1...Parameter RAM, (1!Q...Decoding ROM
0 Agent Yoshihiro Morimoto b (,2) (b) 5, R/EnO14

Claims (1)

[Claims] 1 7'+':','; Perform linear predictive analysis on FJ, and calculate the linear predictive coefficient for every minute V〒hour ll'l+l,
Parameters such as total power and residual power are extracted, and depending on the ratio of total power and residual 1u power, when the residual is small and the value of force/'total force is large, the residual power is A speech analysis encoding method characterized by encoding with a larger bit allocation when the total force value is small.