JPH02282300A - Variable length frame type vocoder - Google Patents

Abstract

PURPOSE:To execute optimization with total distortions and to decrease the total distortions by subjection the representative analysis frame number in a segment to an adaptive control. CONSTITUTION:A frame selector 5 segments the analysis frame from a linear prediction coefft. LPC analyzer 2 and executes the segmental optimum function approximation so as to minimize the total quantity of quantization distortions and frame substitutive distortions. Quantization parameters 102, repeat bits 103 and the total distortions 101 are transmitted with the respective representative analysis frames. A representative frame number determining device 8 compares the total distortions from selectors 5 to 7 and transmits the parameter 202, pit 203 and representative frame number 204 of the min. total distortions. A sound source analyzer 3 executes the window processing of quantization voice signals with the rectangular functions and extracts the data on voiced/ voiceless discrimination and pitch period for each of the fames. The data are passed through a quantizer 4 and is thereby made into sound source data 205. These sets of the data are multiplexed in prescribed format by a multiplexing circuit 9.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a variable length frame type vocoder.

[Conventional technology]

Instead of transmitting all the spectral envelope parameters obtained by analyzing the input audio signal for each analysis frame of a certain length, each of the spectral envelope parameters is selected from M (for example, 20) consecutive analysis frames, and each A variable-length frame type vocoder is known that reduces the amount of transmission information required by transmitting the number of analysis frames represented by these representative analysis frames, so-called repeat bits, along with the spectral envelope parameters of the representative analysis frames. It is being

The representative analysis frame selected for each section is obtained by approximating the distribution of spectral envelope parameters shown by all analysis frames for each section with an optimal approximation function, and this optimal approximation function is an approximation function such as a rectangle, trapezoid, or straight line. are used depending on the purpose of application,
In addition, it is usually a dynamic programming method.
Function settings are made via amming: DP). An example of piecewise optimal function approximation using this DP method is described in detail in Japanese Patent Application Laid-Open No. 62998.

[Problem to be solved by the invention]

Distortion in a variable length frame type vocoder consists of distortion that occurs in the representation of the representative analysis frame itself due to the finite analysis order and quantization number, and distortion that occurs when each analysis frame is replaced with a representative analysis frame. There is distortion. The nature of speech is time-varying, and in particular, due to the time-varying nature of the spectral envelope, if the analysis order, quantization number 2, number of representative analysis frames, etc. are constant, the proportion of each of these distortions in the total amount changes from moment to moment. .

However, in the conventional variable length frame type vocoder described above, since the number of representative analysis filters in a section is fixed, there is a drawback that optimization from the viewpoint of the total amount cannot be expected in most sections.

SUMMARY OF THE INVENTION An object of the present invention is to provide a variable length frame type vocoder that can achieve better optimization than conventional vocoders from the standpoint of total amount.

[Means to solve the problem]

The variable length frame type vocoder of the present invention is equipped with a means for optimizing the number of representative analysis frames within a section within a predetermined range in variable length frame type vocoding using piecewise optimal function approximation.

〔Example〕

Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram showing the analysis side in the first embodiment of the present invention, and FIG. 2 is a block diagram showing the synthesis side.

The analysis side shown in FIG. 1 includes an A/D converter 1, an LPC analyzer 2, a sound source analyzer 3, a quantizer 4, frame selectors 5-7, and a representative frame number determiner 8. , multiplexing circuit 9
It is configured with the following.

The frame selector 5 includes a quantizer 51, a decoder 52, a distortion calculator 53, buffer memories 54 to 57, and a piecewise optimal approximator 58. The frame selectors 6 and 7 are also similar to each other, except that the numerical conditions of the processing performed by each component are not necessarily the same.
is configured the same as.

An input audio signal whose band is limited by a low-pass filter with a cutoff frequency of 3.4 kHz is sampled at a sampling frequency of 8 kHz in the A/D converter 1 and quantized into a predetermined number of bits.

T, PC analyzer 2 performs window processing and analysis of the quantized audio signal input from A, /D converter 1. Each 30m5 minute of quantized audio signal is stored in the built-in memory while 10m5
window processing using a Hamming function] and output as an Oms analysis frame. The quantized audio signal for each analysis frame is subjected to linear predictive analysis, and in this example, the linear predictive coefficient is
ion coefficient (LPC)) is extracted.

The frame selector 5 selects an analysis frame!・・・-20 consecutive words
A piecewise optimal function approximation is performed so that the total number of quantization distortion and frame substitution distortion is minimized, with 200 m5 as one segment and the number of representative analysis frames in the segment being 4 and 17.

The quantizer 51 quantizes the parameters input for each analysis frame from the LPC analyzer 2 using a possible quantization number when the number of representative analysis frames is 4, and buffers the quantized parameters for 20 frames. Write to memory 54. The decoder 52 decodes the child parameters input from the quantizer 51 and writes 20 frames of the decoded parameters to the buffer memory 55. The distortion calculator 53 calculates quantization distortion for each analysis frame using the parameters input from the LPG analyzer 2 and the parameters input from the decoder 52, and calculates 2 of the calculated quantization distortion.
Write one frame of 0 to the buffer memory 56. The K parameters output by the LPG analyzer 2 and before quantization are also written into the buffer memory 57 for 20 frames. The piecewise optimal approximator 58 uses the stored contents of the buffer memories 54 to 57 to perform piecewise optimal function approximation via the DP square method so that the total amount of quantization distortion and frame replacement distortion is minimized.
Extract four representative analysis frames, output the quantized parameters of each extracted representative analysis frame as quantization parameter 02, and repeat the number of analysis frames represented by each representative analysis frame. -103, and the total amount of quantization distortion and frame replacement distortion of each representative analysis frame is output as a total amount 101.

Note that the piecewise optimal approximator 58 is based on the above-mentioned Japanese Patent Application Laid-Open No. 1996-6:
Frame selector 15 described in Publication No. 2-998
(when quantization is performed by pattern matching)
Since it is sufficient to operate with a processing algorithm that is a generalized processing algorithm, the piecewise optimal approximator 5 is
8 is easy to concretely realize.

The frame selector 6 operates in the same way as the frame selector 5 except that the number of representative analysis frames is 5 and the number of quantization parameters is reduced accordingly. The frame selector 7 also operates in the same way as the frame analyzer 5, except that the number of representative analysis frames is six, and the number of quantizations is further reduced accordingly.

The representative frame number determiner 8 compares the total amount input from the frame selectors 5 to 7, selects the quantization parameter and repeat bit input from the frame selector with the minimum total amount, and selects the quantization parameter and repeat bit input from the frame selector with the minimum total amount. 202 and repeat bit 2
Output as 03. Further, the number of representative analysis frames of the selected frame selector is output as the number of representative frames 204.

The sound source analyzer 3 performs window processing on the quantized audio signal input from the A/D converter 1 using a rectangular function at an analysis frame period of 10 m5, and distinguishes between voiced/unvoiced and pitch period for each analysis frame using a known method. Extract the data.

These data are quantized by a quantizer 4, and the sound source data 2
It becomes 05.

The multiplexing circuit 9 multiplexes the input quantization parameter 202, repeat bits 2031 representative frame number 204, and sound source data 205 in a predetermined format determined by the representative frame number 204, and sends the multiplexed data to the synthesis side.

The above is an overall explanation of the operation on the analysis side shown in FIG.

Next, the operation of the main components on the analysis side shown in FIG. 1 when the present invention is applied to a 1200 BPS vocoder will be explained in detail.

FIGS. 5(A), (B), and (C) show the results of the multiplexing circuit 9 when 4°5.6 is selected as the number N of representative analysis frames in each section consisting of M220. The frame structure of one segment of output data is shown. In the case of jN=4 shown in FIG. 5(a), 35 bits are allocated to each parameter of the first to fourth representative frames. or,
Parameter 1 (+, 2+...,
l (8 has 8, 6, 5, 4 degrees 3.3, 3, 3 bits respectively)
s is assigned. 20 sets of parameter vectors [(kA:
i=1. ..., 20), jl, ..., 8] are each quantized to a desired number of bits, and a parameter vector [(kA; 11, ..., 20), j=L...
, 8] to the buffer memory 54 and decoder 52. A buffer memory 54 stores the parameter vector for quantization. The decoder 52 decodes the quantized parameter vector, and decodes the parameter vector C(kA;i21...,'8),j=],.

20] to the buffer memory 55 and distortion calculator 53. Buffer memory 55 stores parameter vectors for decoding. The distortion calculator 53 calculates 2 by the following equation (1).
Quantization distortion of 0 frames δ1°9(j-1...20
) is calculated.

δ3°ゝ=Σw, (lcCiゝ-k A)2(j =
1. , =-, 20) (1) Here wt(i=4
．． ...8) is the spectral sensitivity of the parameter. The calculated 20 quantization distortions are 63°ゝ, (j-1,,20
) is written to the buffer memory 56.

The piecewise optimal approximator 58 uses the parameter vector [(k!';i-1,...
・, 8), j=1. ..., 2o] and buffer memory 5
Parameter vector [(k
(1);1-1. ..., 8L j=1°20] and the quantization distortion δo>, (
j=1. ..., 2o), JP-A-62-998
Using the algorithm disclosed in the publication, when the number of representative frames N=4, the representative frames that can obtain the minimum possible total amount D, 5j', and the sections represented by these representative frames are determined. . Note that in this embodiment, the maximum section length represented by each representative frame is limited to 16 frames. This limiting process is performed to prevent unnecessary allocation of repeat bits, and this level of limiting has little effect on sound quality. Incidentally, if the maximum section length is not limited, the maximum possible section length is 17. The total distortion: j-ball is sent via the output line 101, and the parameters for quantization of each representative frame are sent to the buffer memory 5.
4, and the section length represented by each representative frame is read out via an output line 102 and output as a repeat bit to the representative frame number determiner 8 via an output line 103. The representative frame number determiner 8 calculates min (D, ,
! 7: , C=4.5.6). D4j) and determine the optimal number of representative frames based on the search results. Naori music? 'nl Dmi). are each frame selector 6
．． This is the total distortion calculated by the frame selector 7.

When ff1=4 is optimal, the frame selector 5 performs quantization and selects the determined quantization parameters and repeat bits, which are output to the multiplexing circuit 9. The multiplexing circuit 9 has A=
In the case of 4, a frame having the configuration shown in FIG. 5(A) is output to the assembly and synthesis side. The configuration of the sound source parameters is fixed regardless of ρ and is the same as that in FIG. 5(C). Netsuron M-20
is fixed, and if any three repeat bits are transmitted, there is no need to transmit the other one.

The frame selector 6 is responsible for processing when ρ=5, and the maximum section length is set to 9.

The frame selector 7 is responsible for processing when p=6, and the maximum section length is set to 9.

In FIGS. 5(A), 5(B), and 5(C), the first hit "'F'" is a bit for frame synchronization, FIG. 5(A).

(B) 2nd. 3rd pit pad 00"' and "'01'
Also, the second bit "1" in FIG. 5(C) is a bit that specifies the number of representative frames.

Next, an overall explanation of the operation on the synthesis side will be given using diagrams.

The synthesis side shown in FIG. 2 includes a separation circuit 11 and a decoder 12.
, a parameter interpolator 13, a buffer memory 14,
Decoder 15, pulse generator 16, and noise generator 17
, the switch 18, the LPG synthesis filter 19, and the D/A
The converter 20 is configured to include a converter 20.

The separation circuit 11 separates the multiplexed signal transmitted from the multiplexing circuit 9, and converts the quantization parameter 202 into the decoder 12.
203 to the parameter interpolator 13, the representative frame number 204 to the decoder 12 and the parameter interpolator 13, and the sound source data 205 to the decoder 15.

The decoder 12 decodes the quantization parameter 202 with reference to the number of representative frames 204, converts the decoded parameter into an α parameter, and outputs the α parameter to the parameter interpolator 13. The parameter interpolator 13 refers to the representative frame number 204 and sets the input α parameter to the repeat bit 203.
It is written into the buffer memory 14 repeatedly for the number of analysis frames specified in .

The decoder 15 decodes the sound source data 205, outputs pitch period data to the pulse generator 16, and also controls the switch 18 based on the voiced/unvoiced distinction. When voiced, the pulse generator 16 generates a pulse train at the pitch period specified by the input pitch period data, and this pulse train is selectively output by the switch 18. When there is no voice, the white noise output from the noise generator 17 is selectively output by the switch 18.

The LPG synthesis filter 19 uses the α parameter read from the buffer memory 14 for each analysis frame as a filter coefficient, and is driven by the sound source input from the switch 18 to synthesize a digital audio signal. This digital audio signal is converted into an analog signal by the D/A converter 20 and becomes an output audio signal.

As explained above, in the embodiments shown in FIGS. 1 and 2, the number of representative frames in a section is within the range of 4 to 6 (together with the number of quantization of the K parameter), and the quantization distortion and frame replacement part are Adaptive control is performed so that the total amount of

Note that the embodiment shown in FIGS. 1 and 2 uses parameters as LPC, but α parameter, line spectrum pairs (Line Spectrum Pa1rs; LSP
), etc., can also be used.

The embodiment shown in FIGS. 1 and 2 has been described above.

FIG. 3 is a block diagram showing the analysis side in the second embodiment of the present invention, and FIG. 4 is a block diagram showing the synthesis side.

The analysis side shown in FIG. 3 includes the LPG analyzer 2.2 on the analysis side shown in FIG. Frame selectors 5-71 Representative frame number determiner 8. The multiplexing circuit 9 includes an LPC analyzer 31, frame selectors 32 to 34, and an LPC order determiner 35. Multiplexing circuit 36
It is configured by replacing it with . Frame selectors 32-34
has the same structure as the frame selector 5 on the analysis side shown in FIG. 1, except that the numerical conditions of the processing performed by each component are not necessarily the same, and Reference numbers 61 to 68 are assigned to correspond to reference numbers 51 to 58 assigned to each component of the selector 5. On the synthesis side shown in FIG. 4, the separation circuit 11 on the synthesis side shown in FIG. Decoding RFS, 12, , <Rameter interpolator 13. A buffer memory 14, an LPG synthesis filter 19 are divided into an IL circuit 41 and a decoder 42. Parameter interpolator 43. Buffer memo! J44. ．． It is configured by replacing it with an LPG synthesis filter 45.

The operation of the LPC analyzer 31 is the same as that of the LPC analyzer 2 except that the order of analysis is the 12th order.

The frame selector 32 treats the 12th-order parameters input from the LPC analyzer 31 as 8th-order parameters, and sets the number of representative analysis frames within the 200 m5 section to 6.
Piecewise optimal function approximation is performed so that the total amount of quantization distortion and frame substitute is minimized.

The quantizer 61 ignores the 9th and subsequent 12th order parameters input from the LPC analyzer 31 for each analysis frame and treats them as 8th order parameters for quantization. The distortion calculator 63 calculates the quantization distortion in the quantizer 61 using the unquantized 12th order parameter as a reference. The piecewise optimal approximator 68 uses the distortion that occurs when replacing a non-quantized 12th-order parameter with a parameter for 8th-order decoding (of a representative analysis frame) as a frame substitute. Except for the points mentioned above, the operation of the frame selector 32 is the same as the operation of the frame selector 5.

The frame analyzer 33 operates in the same way as the frame selector 32, except that it treats the input 12th-order parameters as 10th-order parameters and sets the number of representative analysis frames to 5. The frame selector 34 also treats the input 12-order parameters as 12-order as is, and sets the number of representative analysis frames to 4.
does the same thing.

The LPG order determiner 35 operates in the same manner as the representative frame number determiner 8 except that instead of the representative frame number 204, the order of the parameter handled by the selected frame selector is output as the LPC order 206.

Multiplexing circuit 361 Separation circuit 41. Decoder 42° parameter interpolator 43. Multiplexing circuit 91 of buffer memory 44
Separation circuit 11. Decoder 12. Parameter interpolator 13.
The difference from the buffer memory 14 is due to changes based on differences in the formats of parameters handled by these. While the LPC synthesis filter 19 is configured as an 8th order all-pole filter, the LPC synthesis filter 45 is configured as an 8th order all-pole filter.
It is configured as the following all-pole filter.

As can be seen from the above-described differences from the embodiment shown in FIGS. 1 and 2, the embodiment shown in FIGS.
The number of representative analysis frames within the classification is 4 to 6 along with the degree of C.
It is adaptively controlled within the range of individuals.

Note that the embodiments shown in FIGS. 3 and 4 are LPG,! :'
However, for other LPGs, if a converter for converting 12th-order LPC to 8th-order and 10th-order LPC is installed at the input end of the frame selector 32 and 33,
Available.

The embodiment shown in FIGS. 3 and 4 has been described above.

The two embodiments described above optimize the number of representative analysis frames within a section and either the quantization number or the LPG order.
It is also possible to perform optimization by considering all PG orders.

〔Effect of the invention〕

=18 As explained above, the variable length frame type vocoder of the present invention can perform optimization from the viewpoint of total distortion by adaptively controlling the number of representative analysis frames within a section, and has the effect of reducing total distortion. There is.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the analysis side in the first embodiment of the present invention, FIG. 2 is a block diagram also showing the synthesis side, and FIG. 3 is a block diagram showing the analysis side in the second embodiment of the invention. Figure 4 is a block diagram also showing the synthesis side, and Figure 5 is a block diagram showing the synthesis side.
The figure shows the structure of a frame. 1... A/D converter, 2, 31... LPC analyzer, 3... Sound source analyzer, 4... Quantizer, 5~7.32~34 ...Frame analyzer, death...Representative frame number determiner, 9,36...
- Multiplexing circuit, 11.41... Separation circuit, 12.
42...Decoder, 13.43...Parameter interpolator, 14,714...Buffer memory, 15...Decoder, 16...Pulse Generator, 17...Noise generator, 18...Switcher, 19.45...LPC synthesis filter, 20...
...D/A converter, 35...LPC order determiner, 51.61...Quantizer, 52.62...
... Decoder, 53.63 ... Distortion calculator, 5
4-57. 64-67...buffer memory, 58.
68... Piecewise optimal approximator. Agent Patent Attorney Susumu Uchihara

Claims (1)

[Claims] A variable length frame type vocoder using piecewise optimal function approximation, characterized by comprising means for optimizing the number of representative analysis frames within a partition within a predetermined range.