JPH01293398A - Speech encoding system - Google Patents

Abstract

PURPOSE:To improve the quality deterioration of a residue compressing method by providing a 1st power calculating means which calculates the power of a residue waveform as sound source information and a 2nd power calculating means which calculates the power of an encoded sound source waveform. CONSTITUTION:This system has a means 9 which calculates the power of an original residue waveform (or original sound waveform) and a means 19 which calculates the power of a decoded residue waveform (or synthetic speech waveform) and is provided with a means 21 which calculates a coefficient for correcting the amplitude of decoded residue pulses from the outputs of both power calculating means 9 and 19. Therefore, a decrease in residue power which is caused inevitably by the residue compressing method and an accompanying decrease in the power of a synthetic waveform can be compensated. Consequently, the quality of the synthetic speech is improved.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an audio encoding method, and particularly relates to an audio encoding method for converting audio information into 8K.
The present invention relates to a method for improving the quality of encoded audio when compressing it to around bps.

[Conventional technology]

In order to transmit a digitalized audio signal, the audio signal is sampled, quantized, and converted into a binary digital code, and then transmitted using PCM.

On the other hand, when building a communication network using a dedicated digital line, reducing communication costs is a very important issue, and since the amount of information in a voice signal of up to 64 Kbps is too large, it is extremely uneconomical to do so as it is. . Therefore, information compression (that is, low bit rate encoding) of the audio signal for transmission is required.

As an audio encoding method for compressing an audio signal at around 8 Kbps, a method is known in which audio is separated into spectral envelope information and sound source information and each is encoded. Among them, the so-called PARCOR (Partial Auf
This method allows encoding at a low bit rate, but on the other hand, there is a large deterioration in quality. On the other hand, as a method to express the sound source with multiple pulse trains.

Multi-pulse method (see, for example, Ozawa, et al., "Improving the quality of multi-pulse-driven speech coding method" Audio Study Group of the Acoustical Society of Japan Materials 583-78 (1984, 1)), or the residual compression method (as , and others, ``Study of speech synthesis method using residual information,'' Proceedings of the Acoustical Society of Japan 3-1-7 (October 1982), etc.).

As a residual compression method, 1, for example, JP-A-60-150100
The method described in
No. 296398 and JP-A No. 62-194296.

In these methods, the representation of the sound source is N-densified.
The quality is improved compared to the PARCOR method.

[Problem to be solved by the invention]

In the prior art described above, a plurality of pulse trains, which are sound sources, are generated independently for each frame on a fixed basis. Here, a frame is a unit of time in which audio is analyzed.
Usually it is set to about 20 ms.

Assume that the audio waveform has been sampled and converted into a series of sample values X+.

Let the current value be Xt, and then set the two sample values going back to the past as (xt-t)+(i=1.2...+
p). Here, it is assumed that the speech waveform can be approximately predicted from two past samples. The easiest prediction is the glandular shape prediction.

It is assumed that the current value is approximated by multiplying each past sample value by a certain coefficient and adding them together. At this time, the difference between the actual value X at the current point t and the predicted value yt is defined as a prediction error εt. This prediction error fj is called a prediction residual or simply a residual.

The predicted residual waveform of the speech waveform is considered to be the sum of two types of waveforms. One is the so-called error component, whose amplitude is not very large and is close to a random noise waveform, and the other is the error when a pulse due to band vibration is added to the input, which is the error component in prediction. is greatly distorted, resulting in a residual waveform with large amplitude.

This residual component appears repeatedly and periodically, corresponding to the periodicity of the sound source.

Speech is roughly divided into sections with periodicity (speech) and sections where periodicity is not noticeable (no speech), so the predicted residual waveform also shows periodicity in voiced parts. have.

On the other hand, the pulse train generated in the multi-pulse method or the residual compression method can be regarded as an approximation of the residual, and therefore should have periodicity in the voiced portion. Therefore, for example, in the residual compression method, a method disclosed in No. 31-85-5458 is used to extract a predetermined number of representative residual pulses per pitch period, and this is repeated for each pitch period during decoding. Efforts have been made to effectively increase the number of drive sound source pulses. However, in the case of unvoiced sounds, ■f31
-s3-211a or C) 31-86-22 basically means removing small amplitudes from the prediction residual, that is, using only large amplitude residual pulses. Therefore, the effective number of residual pulses that can be used to achieve a predetermined bit rate cannot be increased in the case of voiced speech.

As explained above, the basic principle of the residual compression method is to thin out a part of the residual pulses calculated from the original waveform, so the power of the decoded residual waveform is generally lower than the power of the original residual waveform. It is also normal for it to be smaller. Since the decoded residual pulse is the driving sound source of the synthesis filter, a decrease in the residual power leads to a decrease in the power of the synthesized waveform. Japanese Unexamined Patent Publication No. 61-29
According to the method disclosed in No. 6398, in the case of voiced sounds, a predetermined number of consecutive residual pulses are extracted and transmitted per pitch period, so when the pitch period is long, such as in a male voice, residual pulses that are not transmitted are The ratio of difference pulses increases, and the reduction in power of the decoded residual waveform increases.

This generally means that the power of the synthesized speech is also reduced, and there is a problem in that the loudness value, which is the most basic evaluation measure for determining the quality of speech, is reduced. This is a very basic problem in that the energy of voice is not transmitted accurately.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech encoding method that solves these conventional problems and improves the quality deterioration caused by the residual compression method.

[Means to solve the problem]

In order to achieve the above object, the speech encoding method of the present invention includes:
It has means for calculating the power of the yX residual waveform (or original sound waveform) and means for calculating the power of the decoded residual waveform (or synthesized speech waveform). A feature of the present invention is that it includes means for calculating a coefficient for correcting the amplitude.

[Effect]

The operation of the present invention will be explained using the residual compression method as an example.

In the residual compression method, voiced/unvoiced judgment is performed on the encoding side, and for voiced frames, the prediction residual is thinned out using the method of JP-A-61-296398, and for unvoiced frames, for example, The prediction residual is thinned out and encoded using the method of 194296.

This compressed residual information is transmitted to the decoding side together with auxiliary information such as spectral envelope parameters.

On the other hand, on the decoding side, residuals are decoded from the transmitted compressed residual information. The decoded residual is input to a synthesis filter and one synthesized waveform is output.

The above is an outline of the operation of the residual compression method.Next, the amplitude correction of the decoded residual pulse will be described. This can be broadly classified depending on whether the amplitude correction is performed on the encoder side or the decoder side.
It is further divided into whether it is applied to the residual waveform or the audio waveform. The details will be explained with reference to an embodiment, but the principle will be explained here.

Now, the original audio waveform for one frame is xt(t=1p2.
. If the respective powers are V and U, ■=Σ
...(1) i=1 U=Σ x il 2 ...(
2) i=1, where the power U' of the synthesized speech waveform is obtained by multiplying xI' by the amplitude correction coefficient a.

...(3) becomes. Here, in order to match U' with V, (1)
, From equation (3), the amplitude correction coefficient is:

...(4) becomes. Therefore, if X't is multiplied by the amplitude correction coefficient a calculated by equation (4) to obtain yt=aX t, the powers of the original speech waveform and the synthesized speech waveform after amplitude correction of the corresponding frame will match. Become.

The above description is about the audio waveform, but the same applies to the residual waveform. In addition, in order to correct the amplitude on the encoder side, it is necessary to perform residual decoding (local decoding) of the residual or speech waveform.
decodk), and if this is done on the decoding unit side, it is necessary to transmit the original residual or the power of the original voice.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a first embodiment of the present invention, in which the speech encoding method of the present invention is applied to a speech encoding device (speech GODEC) using a residual compression method.

3(a) is an encoding section, and FIG. 2(b) is a decoding section.

As shown in FIG. 1(a), the encoding unit of the present invention includes a digital audio signal input terminal 1, a buffer memory 2 for storing the signal, a linear prediction circuit 4 for performing linear prediction, and a spectral envelope parameter. 5, a residual thinning circuit 8 that extracts a representative residual from the residual waveform 7, and a first power calculation that calculates the power of the residual waveform 7 for one frame. It includes a circuit 9 and a quantization hakama circuit 12. Further, as shown in FIG. 1(b), the decoding section of the present invention includes a decoding and inverse quantization circuit 15 that separates the input signal into 381 parameters, and a buffer that stores the decoded spectral envelope parameter 5'. A memory 17, a residual interpolation circuit 16 that reproduces the sound source pulse from the decoded compressed residual information 10', a second power calculation circuit 19 that calculates the power of the sound source pulse, that is, the decoded residual 18, and its output. A correction coefficient calculation circuit 21 which receives the decoded residual power 20 and the decoded original residual power 11' as input, a multiplier 23 which multiplies the decoded residual 18 by the correction coefficient 22 which is the output thereof, and a synthesis filter 1. 25 and an output terminal 27 for synthesized speech 26.

In FIG. 1(,), at the time of encoding, the digitized audio signal is sent to a buffer memory 2, divided into frames, and is processed by a well-known linear prediction circuit 4 using parameters representing the spectral envelope (for example, polarization). (autocorrelation coefficient) 5.

Next, an inverse filter 6 is configured using this spectral envelope parameter 5 as a coefficient, and the audio signal 3 is inputted to this to obtain a residual waveform signal 7.

The residual waveform signal 7 is input to a residual thinning circuit 8, and compressed residual information 1o is obtained as its output. Residual thinning circuit 8
Although various configurations can be considered, here,
It is assumed that the residual thinning method disclosed in JP-A No. 96398 and JP-A No. 62-194296 is used. Both are 1
Multiple residual pulse trains called representative residuals are extracted from the residual waveform signal 7 for one frame, and information regarding their amplitudes (the maximum amplitude and the amplitude of each residual pulse normalized by it) and positions are compressed. It is output as residual error information 10.

The first power calculation circuit 9 is a circuit added according to the present invention, and calculates the residual power 11 for the residual waveform signal 7 or 61 frames by the following equation.

■=Σ Z+"...
・(5) i=1 Here, Z indicates the amplitude of the residual pulse at address i within the frame, and Ω indicates the number of residual pulses within one frame.
6 where ■ is the value of the residual power 11. The quantization encoding circuit 12 receives the spectral envelope parameter 5, the compressed residual information 10, and the residual power 11, quantizes it to a predetermined number of bits, and converts it to a predetermined number of bits. The digital data 13, which is the result of conversion into the format, is sent to the digital line 14.8 Figure 1(b)
At the time of decoding, when the digital data 13 received from the digital line 14 is input to the decoding and inverse quantization circuit 15, the three types of parameters shown in (a) (spectral envelope parameter 5', compressed residual information 10 ', [residual power 11')]. Among the above parameters, the spectral envelope parameter 5' is temporarily stored in the buffer memory 17, and the compressed residual information 10' is stored in the residual interpolation circuit 1.
6, and a decoded residual 18 is output. As the residual interpolation circuit 16, a method disclosed in Japanese Patent Laid-Open No. 61-296398 and Japanese Patent Laid-open No. 62-194296 is used.
A decoded residual waveform 18 for one frame serving as a driving sound source is output. The second power calculation circuit 19 is a circuit added according to the present invention, and receives the decoded residual waveform 18 as input,
The decoding residual power 20 is calculated using the following equation.

Ω Here, Zi' is the amplitude of the decoded residual pulse at address i within the frame, and is a value obtained by quantizing Z in equation (5). Q is the same as equation (5), and U is the value of the decoding residual power 20. The correction coefficient calculation circuit 21 is a circuit added according to the present invention, and uses the decoded original residual power 11' and the decoded residual power 20 obtained by equation (6) as input, and calculates the correction coefficient 22 by the following equation. do.

・=Gishifu τ ...(7) V here
' is the value of the decoded original residual power 11', which is obtained by quantizing the value V calculated by equation (5). a is the value of the correction coefficient 22, which is multiplied by the decoded residual 18 by the multiplier 23 to obtain the decoded residual waveform 24 whose amplitude has been corrected.

After correcting the delay in the above processing, the output of the buffer memory 17 is used as the coefficient of the synthesis filter 25. By manually inputting the amplitude-corrected decoding residual 24 to the synthesis filter 25, a synthesized speech 26 is obtained as an output, and is connected to an output terminal 27.

As explained above, (1) according to the present invention, it is possible to correct the decrease in residual power caused by thinning out residuals by the residual compression method, (1) the power of the synthesized speech becomes close to the power of the original speech, and (1) The quality of synthesized speech can be improved.

Next, a second embodiment of the present invention will be described.

Figure 2 shows an audio encoding system using the residual compression method as the audio encoding method of the present invention! ! ! FIG. 3 is a block configuration diagram of a second case applied to (audio CODEC).

(a) is an encoding section, and (b) is a decoding section.

The second embodiment is different from the first embodiment in that the inputs to the first power calculation circuit 28 and the second power calculation circuit 3o are the original audio signal 3 and the synthesized audio 31 for one frame, respectively. Different from the example. In the second embodiment, which differs in that the synthesized speech 31 is multiplied by a correction coefficient 22, the basic principle is to apply amplitude correction directly to the synthesized waveform. below,
The explanation will focus on the differences from the first embodiment.

In FIG. 2(a), first, the original audio signal 3 for one frame is input to the first power calculation circuit 28, and the original audio power 29 is calculated using equation (1) (the value is set to V), The signal is sent to the quantization encoding circuit 12.

In the same figure (b), decoded compressed residual information 10
' is input to a residual interpolation circuit 16, and a decoded residual 18 is output. By inputting this decoded residual 18 to a synthesis filter 25 whose coefficient is the decoded spectral envelope parameter 5' passed through the buffer memory 17, a synthesized speech 31 is output. In the second power calculation circuit 30 (2)
The power 32 of the synthesized speech is calculated according to the formula (the value is set to U).

The correction coefficient calculation circuit 21 receives the original audio power 29' that is input to the decoding and dequantization circuit 15 via the digital line 14 and is separated and decoded (the value is V' as a result of quantization); The aforementioned synthesized speech power 32 is input, and the correction coefficient 22 is calculated using equation (7). This correction coefficient 22 is multiplied by the synthesized speech waveform 31 in a multiplier 23 to obtain an amplitude-corrected synthesized speech 33, which is connected to the output terminal 27.

As explained above, according to the present invention, the decrease in the power of synthesized speech due to the decrease in residual power caused by thinning out the residual by the residual compression method is corrected, and the power of the synthesized speech matches the power of the original speech. Therefore, the quality of synthesized speech can be improved.

Next, a third embodiment of the present invention will be described.

In the first and second embodiments described above, it was necessary to transmit the output result of the first power calculation circuit 9 or 28, but in this embodiment and the fourth embodiment described below, the code section The feature is that by providing a local decoder on the side, there is no need to transmit power.

The third embodiment will be described below first.

FIG. 3 is a block diagram of the third case in which the audio encoding method of the present invention is applied to an audio encoding device (audio C0DEC) using the residual compression method, in which (a) shows the encoding section. , and (b) is the decoding section.

The feature of this embodiment is that in the encoding section shown in FIG. 2, a residual interpolation circuit 34 is provided after the residual thinning circuit 8, and this acts as a local decoder for the residual waveform. The residual interpolation circuit 34 operates in the same manner as the residual interpolation circuit 16 shown in FIG. Output. Here, JP-A No. 61-296398 and JP-A No. 62-19
This method is based on the method disclosed in No. 4296. The first power calculation circuit 9 receives the residual waveform 7 as an input and calculates the residual power 11 (the value thereof is V) using equation (5). The second power calculation circuit 19 inputs the local decoding residual 35 and (
By inputting the two column difference powers 11 and 20, which are decoded residual power 20 (its value is U) according to the formula 6), the correction coefficient calculation circuit 21 calculates the correction coefficient 22 according to the following formula.

a=Pf〒 (8) where a is the value of the correction coefficient 22. Next, this correction coefficient 22 is multiplied by the compressed residual information 10 by the multiplier 23, but the compressed residual information 10 is divided into position information and amplitude information as described above, and the amplitude information is Maximum absolute value amplitude ZMX
and normalized amplitude. In other words, if the amplitude of the representative residual immediately after residual thinning is Zl', then Zs' =
The relationship is Zt' / ZMX - (9), and the amplitude information actually transmitted is ZMX and (zt
), therefore, the correction coefficient 22 is the maximum amplitude ZM
Multiplying only X is equivalent to correcting the amplitudes of all representative residuals during decoding. In this way, the maximum amplitude ZM
Compressed residual information 36 obtained by multiplying X by the correction coefficient 22 is sent to the quantization encoding circuit 12.

In the decoding section shown in FIG. 4B, the decoding and dequantization circuit 15 separates the signals. The amplitude-corrected compressed residual is input to the residual interpolation circuit 16. The residual amplitude information is restored by the inverse operation of equation (9).

That is, Z1' = Zt'・ZMX' ・・
・(10) Here, ZMX' is the amplitude-corrected residual amplitude maximum value ZMX' = a − ZMX
= (11), and from this and the position information, the decoding residual 39
is output. The operation of residual interpolation 16 is described in Japanese Patent Application Laid-open No. 61-29.
6398 and Japanese Patent Laid-Open No. 62-194296. The decoding residual 39 is stored in the buffer memory 17
is input to a synthesis filter 25 whose coefficient is the spectral envelope parameter 5' passed through
is sent to the output terminal 27.

According to this embodiment, the residual amplitude can be corrected in the coding section, and the original residual power and the decoded residual power can be matched without transmitting power information. As a result, the voice power is transmitted almost accurately, and the quality of the synthesized voice can be improved.

Next, a fourth embodiment of the present invention will be described.

FIG. 4 is a fourth block diagram configuration diagram in which the audio encoding method of the present invention is applied to an audio encoding device (audio C0DEC) using the residual compression method, in which (a) is the encoding section; ,(
b) is a decoding unit.

The feature of this embodiment is that, in the encoding section of FIG. It is located at the points constituting (within the broken line in Figure (a)).

Compressed residual information 10, which is the output of the residual thinning circuit 8, is input to the residual interpolation circuit 34, where it is locally decoded and a decoded residual 35 is output. This decoded residual 35 is input to a synthesis filter 37 that uses the spectral envelope parameter 5 as a coefficient, and decoded speech 38 is locally decoded.

Similar to the second embodiment, the first power calculation circuit 28 receives the audio waveform signal 3 and outputs the original audio power 29 (the value thereof is assumed to be V). The second power calculation circuit 30 receives the locally decoded speech 38 as input, and calculates the synthesized speech power 32 using equation (2) (the value thereof is assumed to be U). Correction coefficient calculation circuit 21
The above two types of powers 29 and 32 are input to , and the correction coefficient 22 is calculated according to equation (8). Hereinafter, the execution of amplitude correction on the compressed residual information 10 is exactly the same as in the third embodiment. The reason why the amplitude of the compressed residual information 10 is corrected based on the audio power is to take advantage of the property that the magnification of the residual power and the magnification of the audio power are proportional. When the residual information 36 is used, the voice power matches the original voice and the synthesized voice.

The operation of the decoding section in FIG. 2B is exactly the same as in the third embodiment, so a description thereof will be omitted.

As explained above, according to this embodiment, there is no need to transmit power information and the power of the synthesized speech can be made to match the original speech, so the decrease in the synthesized speech power that occurs due to the residual compression method can be avoided. This can improve the quality of synthesized speech.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to compensate for the decrease in residual power that inevitably occurs due to the residual compression method and the accompanying decrease in the power of the synthesized waveform, thereby improving the quality of synthesized speech. It is possible to improve. In particular, in the case of a configuration in which a local decoder is provided on the code section side, there is no need for transmission parameters newly added to the conventional method, and an increase in the bit rate can be suppressed.

[Brief explanation of the drawing]

(a) and (b) of FIGS. 1 to 4 are the first embodiments of the present invention, respectively.
FIG. 12 is a block diagram of a speech encoding device according to a fourth embodiment of the present invention, showing an encoding section and a decoding section. 4...Linear prediction circuit, 6...Inverse filter, 8...
Residual thinning circuit, 9, 28... first power calculation circuit,
16°34...Residual interpolation circuit, 19.30...Second
power calculation circuit, 21... correction coefficient calculation circuit, 23.
...Multiplier. 25.37...Synthesis filter. Figure 1 (α) Figure 2 (Good

Claims (1)

[Claims] 1. Analyzing the audio signal at fixed time intervals (frames),
In a speech encoding method that separates spectral envelope information and sound source information and uses multiple pulses as a driving sound source,
a first power calculation means for calculating the power of the residual waveform that is the sound source information; and a second power calculation means for calculating the power of the decoded sound source waveform; and a correction coefficient calculating means for correcting the amplitude of the sound source waveform from the result of the sound encoding method. 2. Analyze the audio signal at fixed time intervals (frames),
In a speech encoding method that separates spectral envelope information and sound source information and uses multiple pulses as a driving sound source,
The former includes a first power calculation means for calculating the power of the audio signal waveform, and a second power calculation means for calculating the power of the decoded audio waveform synthesized with the decoded sound source waveform as a driving sound source, and a correction coefficient calculation means for correcting the amplitude of the decoded speech waveform from the results of the first and second power calculation means. 3. Analyze the audio signal at fixed time intervals (frames),
In a speech encoding method that separates spectral envelope information and sound source information and uses multiple pulses as a driving sound source,
a first power calculation means for calculating the power of the residual waveform which is the sound source information; and a second power calculation means for calculating the power of the decoded sound source waveform, the means comprising a local decoding means for the sound source waveform. and a second power calculation means for calculating the power of the decoded excitation waveform that is the output of the local decoding means,
and a second power calculating means, the encoding unit further comprising a correction coefficient calculating means for correcting the amplitude of the sound source waveform from the results of the second power calculating means. 4. Analyze the audio signal at fixed time intervals (frames),
In a speech encoding method that separates spectral envelope information and sound source information and uses multiple pulses as a driving sound source,
The encoding unit includes a first power calculation means for calculating the power of the original audio signal waveform, a local decoder for the sound source waveform, and a local decoder for the audio waveform using the output of the local decoder for the sound source waveform. The encoding unit includes second power calculation means for calculating the power of decoded speech that is the output of the local decoder of the speech waveform, and from the results of the first and second power calculation means,
A speech encoding method, characterized in that an encoding section is provided with a correction coefficient calculating means for correcting the amplitude of the sound source waveform. 5. The speech encoding method according to claim 1 or 3, characterized in that the power of the original residual and the power of the decoded residual are made equal to each other. 6. The speech encoding method according to claim 2 or 4, characterized in that the power of the original speech and the power of the synthesized speech are matched. 7. A multimedia multiplexing device using the audio encoding method according to claims 1 to 6. 8. A mobile communication system using the voice encoding method according to claims 1 to 6. 9. A voice mail and voice response device using the voice encoding method according to claims 1 to 6.