JPS62997A - Multi-buss endoder/decoder - 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] [Industrial application field] The present invention relates to a multipulse encoding/decoding device.

[Conventional technology]

The sound source information of the input audio signal is represented by a plurality of impulse sequences, so-called multipulse sequences, and is transmitted along with information regarding the spectral envelope from the encoding side (analysis side) to the decoding side (synthesizing side) to reproduce the input audio signal. Multipulse encoding/decoding devices are well known these days.

The encoding of a multi-pulse train in such a multi-pulse encoding/decoding device is performed by increasing the amplitude of the multi-pulse train by 2 to 5.
A quantized word with a finite word length on the order of bits is transmitted to the decoding side along with spectral envelope information.

[Problem that the invention seeks to solve]

This type of conventional multi-pulse encoding/decoding apparatus has a problem in that quantization noise generated during quantization of a multi-pulse train degrades the synthesized sound quality on the decoding side.

The reason for this is not only that the power of the quantization noise itself cannot be ignored, but also that the multipulse train and the quantization noise are not uncorrelated. For example, in voiced speech, the multi-pulse train to be analyzed has pitch periodicity, and the quantization noise generated as a result of quantizing this multi-pulse train also has pitch periodicity.

An object of the present invention is to provide a multipulse encoding/decoding device that can significantly improve the deterioration of synthesized sound quality by whitening quantization noise.

[Means for solving problems]

The device of the present invention is a multipulse encoding/decoding device, and includes an encoding side that is equipped with a noise addition means that adds noise that is uncorrelated with the extracted multipulse sequence to the extracted multipulse sequence; and a decoding side, which is provided with a noise subtraction means that generates noise in synchronization with the multipulse train and subtracts this noise from the decoded multipulse train.

〔Example〕

Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of a multipulse encoding/decoding device according to the present invention.

The configuration of the embodiment shown in FIG. 1 includes an encoding side 1 and a decoding side 2, and the encoding side 1 further includes a multipulse analyzer 11, an LPG analyzer 12, and so on. Quantizer 13. Adder 1
4. Noise generator 15. It is configured to include a quantizer 16, an encoder 17, and a multiplexer 18. Further, the decoding side 2 includes a demultiplexer 21. Decoders 22, 23
, 24 , subtractor 25 , speech synthesis filter 26 , D
In addition to the /A converter 27 and the LPF 2g, the encoder includes the same noise generator 15 as the encoding side.

On the encoding side 1, the input audio signal input via the input line 1001 is sent to the multipulse analyzer 11 and the LPC analyzer 1.
2.

The multi-pulse analyzer 11 performs a multi-pulse search based on a known multi-pulse search method, so-called correlation region evaluation in this embodiment, to obtain a multi-pulse train of the input audio signal.

This correlation domain evaluation is performed by converting the input audio signal S■ into a transfer function W■
The auditory weighting is the output S■*W obtained through the filter.
■, input audio signal to LPC (Linear Pred)
Calculate the cross-correlation function between the vocal tract filter coefficient H■ obtained by analyzing the input coefficient (linear prediction coefficient) and the impulse response of H-cut *W■ obtained through the weighting filter W■ for the analysis frame, In addition, the autocorrelation of the impulse response of H■*W■] is evaluated by determining the multipulse train while searching for the one that minimizes the residual waveform component obtained by subtracting this autocorrelation function from the cross-correlation function. It is a method.

In addition, as mentioned above, 7tZ is exp (jλ) and λ-2xTΔ
f1ΔT is the analysis frame sampling period, f is the frequency, and * is the convolution integral.

The LPC analyzer 12 extracts LPC coefficients of a predetermined order for each analysis frame of the input audio signal, and supplies them to the multipulse analyzer 11 and the quantizer 13 . The LPC coefficients provided to the multipulse analyzer 11 are used for multipulse search based on the above-mentioned correlation region evaluation, and the LPC data provided to the quantizer 13 is quantized with predetermined content and then sent to the multiplexer 18. is supplied as spectral envelope information. In the analysis of LPC coefficient data by this LPC divider 12, the input audio signal is set at a predetermined sampling frequency, in this example, gk
Hz, multiplication with a predetermined window function is performed for each analysis frame, and an LPC coefficient of a predetermined order is extracted for each analysis frame period and this is supplied to the multipulse analyzer 11. is being carried out. Therefore, the searched multi-pulses are also gkH
One of the sampling points of z will be determined as the time position.

The next multi-pulse train searched in this way is supplied to the adder 14, where it is added to the noise supplied from the noise generator 15.

The noise supplied from the noise generator 12 is selected to be uncorrelated with the multi-pulse train, and is applied to the multi-pulse at the sampling point where the multi-pulse is being searched among the 8 kHz sampling points.

In the case of this embodiment, the noise is generated using a 15-bit M-sequence code, and while this code is stepped by an 8kHz clock signal, the noise is generated using an 8kHz clock signal set with a multi-pulse.
z Added to each of the multipulses at the sampling time position. This addition was carried out with the aim of making the quantization noise have a random probability of occurrence by superimposing the noise generated using a code such as an M-sequence code whose probability of occurrence is random and then quantizing it. It is something to do.

The noise generator 15 further includes a noise generator 1 on the decoding side.
Synchronization data for noise generation synchronization for No. 5 is also generated. This synchronization data may be performed using the same method as random number synchronization in ordinary digital secret signals, such as using initial data or using a synchronization code set in advance. After synchronization, the initial data output from the noise generator 15 is encoded by the encoder 17 and then supplied to the multiplexer 18.

The quantizer 16 receives the output of the adder 14, quantizes it by a predetermined number of bits, and supplies it to the multiplexer 18 as multi-pulse data. The multi-pulse data obtained by K in this way is generated using the original multi-pulse train and the M-sequence code, and contains quantization noise during quantization in addition to the superposition of nine noises.Moreover, this quantization noise is It is white noise.

The multiplexer 17 receives this multipulse data.

The above-mentioned synchronized data and LPC coefficient data are multiplexed in a predetermined format and then supplied to the decoding side 2 via the transmission line 1002.

After demultiplexing the multiplexed signal thus supplied on the decoding side 2, the multipulse data is sent to the decoder 22.
The synchronization data is supplied to a decoder 23, and the LPC coefficient data is supplied to a decoder 24 for decoding.

The subtracter 25 receives the decoding output supplied from the decoder 22,
That is, the noise generated by using the M-sequence code that advances at 8 kHz is removed by subtracting only the noise weighted on the analysis side from the multipulse train.

This subtraction is performed from the output of the decoder 22, which generates noise in synchronization with the encoding side under the control of synchronization data that the noise generator 15 on the decoding side 2 receives from the noise generator 15 on the encoding side l. This is done with the aim of reducing

By performing such subtraction, the subtracter 25 applies a method to whiten the quantization noise, removes the whitening noise from the multi-pulse train, outputs a multi-pulse train in which only the whitened quantization noise remains, and uses this as a driving sound source to generate audio. It is supplied to the synthesis filter 26.

On the other hand, the decoder 24 decodes the LPC coefficients and supplies them to the speech synthesis filter 26.

The speech synthesis filter 26 is configured as an all-pole digital filter, and is driven by the driving sound source using the multi-pulse train using the LPC coefficients supplied from the decoder 24 as filter coefficients, and reproduces the digital input speech signal. Further, this digital input audio signal is converted into an analog quantity by a D/A converter 27, unnecessary frequency components are removed by an LPP 28, and the signal is output to an output line 2001.

In this way, the quantization noise of the multi-pulse train can be whitened and the reproduced sound quality can be significantly improved.

Next, the noise generator 15 will be explained in detail using the drawings. FIG. 2 is a block diagram for explaining the noise generator 15. The noise generator 15 shown in FIG.
-1 to 15. It is configured to include an exclusive OR 52. The sign of 1 bit is inputted to the exclusive OR 52 from the shift registers 51-14 and 51-15. The exclusive OR 52 calculates the exclusive OR of these two codes and outputs it to the shift register 51-1. shift register 5
1-1 to 15 step in synchronization with the clock signal supplied from the terminal 55. Such a combination of a shift register and an exclusive OR is known as an M-sequence. The number of clock signals supplied through terminal 55 is the same number of pulses as the number of audio samples corresponding to the length of one frame, in this embodiment the number of pulses per frame. The outputs of the shift registers 51-1 to 51-9 are output to noise output terminals 56-1 to 9-9.
>Supplied. Note that 56-1 to 8 are codes each having a weight of 2 to 2, and 56-9 is a code having a weight of -2. Therefore, the value of the noise output from the noise output terminals 56-1 to 56-9 is -28 to 2. '-1. The distribution range of this noise is determined and proportional to the step size of multi-pulse amplitude quantization performed by the quantizer 16. Now, the noise generator 15 on the encoding side and the noise generator 15 on the decoding side must step in synchronization. In this embodiment, synchronization is established by transferring initial data regarding the step position of the M sequence from the encoding side to the decoding side for each frame. The noise generator on the encoding side generates K 1 per frame.
After being supplied with 60 clocks and incrementing, the contents of the shift register 51-1-15 are transferred to the initial value output terminals 57-1 to 1.
5 to the encoder 17. The encoder 17 uses the supplied 15-bit data as synchronization data.
l. Handle without processing. The contents of the shift register 51-1-15 included in the noise generator 15 on the encoding side are transferred to the initial state of the noise generator 15 on the decoding side via the encoder 17, multiplexer 18, demultiplexer 21, and decoder 23. The signal is supplied to the value terminal 53-1-15. The noise generator 15 on the decoding side is inputted into the shift registers 51-1 to 51-15 by a preset signal supplied from the preset control terminal 54.

Note that this preset signal is supplied for each frame prior to the occurrence of noise.

In the embodiment shown in FIG. 2, the 15th order M sequence is used as a noise generator because its period is 2 as is known.
-1-32767, this is because even if noise is generated in response to the audio sample, if the sampling rate is gkHz, the noise is guaranteed to be 32768/8X10"#4 seconds. This time length is It is so long that it cannot be remembered aurally.

However, in this embodiment, a 15-bit initial value is transmitted for each frame, but this can be reduced by the following method.

The first method utilizes the fact that the number of multipulses is generally 1/8 or less than the number of audio samples in one frame. In the embodiment shown in FIG. 2, the noise generator 15 advances by D 160 (8000 samples per second with a 7-frame period of 20 m5 ec) samples per frame. This is made to advance by the number of multipulses per frame. Set the number of multipulses per frame to 2.
If it is set to 0, the noise consumption will also be reduced to 20/16=1/8. Therefore, if the period of the noise is the same, a 12th order M sequence (period 2-1=4095) can be used. In this case, the number of bits required for initial value transmission is 12 bits per frame.
It becomes a bit.

The second method is to generate noise using a plurality of random numbers with different periods or a combination of noises. In this method, the period of one random number is made to match the frame period, etc. For example, if the frame period is 160 (8kHz samples, frame length 2)
0 m58e), and the period of one noise source is 160
, the period of the other noise source is 511, then the period of the noise created by combining the two noises is the least common multiple of 160 and 511, that is, 81,760. In this case, period 1
By synchronizing the 60 noise sources with the frame, it is easy to synchronize the encoding and decoding sides. Also, the period 511
The noise source can be composed of a 9th order M sequence. in this case,
The number of bits required for initial value transmission is per frame9b
It becomes it. In thermal theory, it is also possible to implement a ratio method that combines the two methods described above. Audio sampling rate gkHz,
If the frame period is 20 m5ec (the number of samples per frame is 160) and the number of multipulses per frame is 20, then the period of one noise source is 22,
The period of other noise sources can be achieved by the 8th order M-sequence.
5, and by further advancing the noise source by 22 frames, the effective period is (least common multiple of 22 and 255) ÷ (8000 is D(22-20) per frame)
= 2 steps without using noise. still,
In this case, the number of initial value transmission bits required is 8 bits.
becomes.

Next, another embodiment of the noise generator 15 will be described using the drawings. FIG. 3 is a block diagram showing another configuration of the noise generator 15. The noise generator 15 shown in FIG.
1. ROM62, Power? 63 and a full adder 64. The 8th order M sequence "61" and the counter 63 step in synchronization with the clock supplied from the clock input terminal 65. This clock is supplied with 22 pulses per frame. The counter 63 is a 5-bit binary counter. It is preset to '0#' every frame.A signal to control this presetting is supplied from the terminal 66.The output of the counter 63 is used as an address signal for the ROM 62.The ROM 62 is preset.

A g-bit random number is written in the range of addresses θ to 21. This kind of random number is created using a random number table.几0
The g bit output of M62 is provided to full adder 64. The 8-bit output of the 8th order M sequence 61 is supplied to the full adder 64 and initial value output terminals 68-1 to 68-8. Full adder 6
4 is each f3 supplied from the ROM 62 and the 8th order M series 61
Add bit data and output noise output terminals 69-1 to 69-9.
Output to. The method for establishing synchronization between the noise generators 15 on the encoding side and the decoding side is basically the same as the method shown in FIG. 2], and a detailed explanation will be omitted. In FIG. 3, the initial values supplied to the decoding side are at terminals 70-1 to 70-8.
Data preset control is performed via terminal 67.

〔Effect of the invention〕

As explained above, according to the present invention, in a multipulse encoding/decoding device, deterioration in synthesized sound quality can be significantly improved by providing means for whitening multipath sequence (Q quantization noise). This has the advantage that it is possible to realize a multi-pulse encoding/decoding device that can perform

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of an encoding/decoding apparatus according to the present invention, and FIGS. 2 and 3 are block diagrams of a noise generator 15. l...Encoding side, 2...Decoding side, 1
1...Multi-pulse searcher, 12...
LPC analyzer, 13...Quantizer, 14...
...Adder, 15...Noise generator, 16...
... Quantizer, 17 ... Encoder, 18.
...Multiplexer, 21 ...Demultiplexer, 22.23.24 ...Decoder, 2
5...Subtractor, 26...Speech synthesis filter, 27...D/A converter, 28...
...LPF, 51-1-15 Old...Shift register, 52...Exclusive OR, 61...8
Next M series, 62...ROM, 63...
Counter, 64... Full adder 〇 Agent Patent attorney Susumu Uchihara r3 diagram

Claims (1)

[Claims] In a multi-pulse encoding/decoding device, an encoding side is provided with a noise addition means for adding noise uncorrelated with the extracted multi-pulse sequence to the extracted multi-pulse sequence, and the noise is generated in synchronization with the encoding side. and a decoding side equipped with noise subtraction means for subtracting this noise from the decoded multipulse train.