JPH04249300A - Audio code/decoding method and device - 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]

0001

[Industrial Application Field] The present invention is an audio encoding/decoding method and apparatus for highly efficiently encoding an analog audio signal, sending it to a transmission path, decoding it on the receiving side, and decoding and reproducing the analog audio signal. It is related to.

0002

2. Description of the Related Art FIG. 3 is a block diagram of a conventional evenly spaced pulse-driven speech codec using long-term prediction, in which (A) is a speech encoding device and (B) is a speech decoding device. This method has a coding speed of 13 Kbps (bits per second) and is a voice coding/decoding method adopted in the Pan-European Digital Mobile Telephony (GSM) system. Below, the encoding speed is 64
A method for compressing from Kbps to 13 Kbps will be explained.

In FIG. 3A, an input audio signal (64 Kbps) 8-bit quantized with 8 kHz sampling is shown.
) is calculated by the short-term prediction analyzer 11 for each frame (160
Short-term predictive analysis (also referred to as linear predictive analysis) is applied to the sample (20 msec). That is, it extracts and outputs the spectral envelope information Pa from the input audio signal, and also generates and outputs the short-term prediction residual signal a (160 samples), which is a signal from which the spectral envelope component has been removed. Next, the short-term prediction residual signal a is divided into four subframes (40 samples) by the long-term prediction analyzer 12, and pitch information Pb is extracted and output for each subframe, and the signal is further removed from the pitch component. The long-term prediction residual signal b(
40 samples) is generated and output. This long-term prediction residual signal b (40 samples) is band-limited to ⅓ by an LPF (low pass filter) 13 to obtain a signal c. The signal c is down-sampled to 1/3 by a switch S having three grids, but at this time the grid selector 14
The grid signal sequence (13 samples) with the maximum power is selected, and these are output as equally spaced pulse information Pc. The above Pa, Pb and Pc are encoder 1
5 and is sent to the receiving side as a digital string encoded and multiplexed. The bit allocation per frame for each parameter at this time is as shown in Table 1. However, as the equally spaced pulse information Pc, Part 1
Maximum value among 3 samples and 13 normalized by it
It consists of a sample and position information (grid number). Since one frame is 20 msec, the encoding speed is 13 Kbps.

[Table 1]

In FIG. 3(B), the received digital string is separated into equally spaced pulse information Pd, pitch information Pe, and spectrum envelope information Pf by a separation circuit 21, and then is separated into equally spaced pulse information Pd, pitch information Pe, and spectrum envelope information Pf by a long-term prediction residual signal regenerator 22. A long-term prediction residual signal d is reproduced from the pulse information Pd. Here, the sent evenly spaced pulses are rearranged to the original grid positions, and 0's are inserted at sample points where no equally spaced pulses exist. Next, the long-term prediction synthesizer 23 adds pitch information Pe to the long-term prediction residual signal d, and reproduces the short-term prediction residual signal e. Next, the short-term prediction synthesizer 24 adds spectrum envelope information Pf to the short-term prediction residual signal e and outputs a reproduced audio signal.

[0005]

Problems with the above-mentioned conventional method will be explained using FIG. 4. FIG. 4(A) is the spectrum of the long-term prediction residual signal b. If this is filtered by the ideal LPF of FIG. 4(B), a signal as shown in FIG. 4(C) is obtained, which is 1/ No aliasing distortion occurs even with downsampling of 3. However, filtering in the conventional method is achieved by convolving the impulse response of the ideal LPF with the long-term prediction residual signal sequence in the time domain, and the impulse response, which should be infinite, is processed using only 11 samples. Figure 4 (D
), and the signal filtered by this spectrum becomes as shown in FIG. 4(E). This is 1/
If downsampled to 3, aliasing distortion (shaded area) will occur as shown in FIG. 4(F). This affects the reproduced sound, causing a phenomenon in which the voice uttered by one person sounds as if it were uttered by two people. Furthermore, since the aliasing distortion becomes large, it is impossible to thin out the samples of the long-term prediction residual signal any more, and it is difficult to reduce the encoding speed lower than this.

To summarize the above, the drawbacks of the conventional system are as follows. (1) Folding distortion occurs due to imperfections in the LPF, and the quality of reproduced sound deteriorates. (2) It is difficult to reduce the bit rate.

SUMMARY OF THE INVENTION An object of the present invention is to provide a method and apparatus for encoding and decoding speech at a lower encoding speed while reducing the adverse effect on quality caused by aliasing, which is a drawback of conventional methods.

[0008]

[Means for Solving the Problems] Fig. 1 is a block diagram of a speech code/decoder showing an embodiment of the present invention, in which (A) is a speech coder and (B) is a speech decoder.

In FIG. 1A, an input audio signal (64 Kbps) 8-bit quantized with 8 kHz sampling is shown.
) is subjected to short-term predictive analysis for each frame (160 samples: 20 msec) by the short-term predictive analyzer 31. That is, it extracts and outputs the spectral envelope information Pg from the input audio signal, and also generates and outputs the short-term prediction residual signal g (160 samples), which is a signal from which the spectral envelope component has been removed. Next, this short-term prediction residual signal g is divided into four subframes (40 samples) by the long-term prediction analyzer 32, and pitch information Ph is extracted and output for each subframe, and the pitch component is further removed. A long-term prediction residual signal h (40 samples) is generated and output. This long-term prediction residual signal h (40 samples) is transformed into the frequency domain by a discrete cosine transform (DCT) unit 33 and outputs a DCT coefficient i. The DCT conversion formula will be described later. Next, the DCT coefficient i (40 samples) is decimated by a decimator 34 to be represented by 7 samples. FIG. 2 is an explanatory diagram of the thinning method. Figure 2
(A) is the result of DCT transformation of the long-term prediction residual signal. As shown in the same figure (B), first, 1.33 [kHz
] or more are thinned out. This is the LP for the conventional method.
It has the same effect as limiting the band to 1/3 by F, but
Since the component is canceled in the frequency domain, the frequency is 1.33 [kHz
] No components remain and aliasing distortion is reduced. Furthermore, in order to further reduce the encoding speed, the solid line or the dotted line in the figure, whichever has the higher power, is selected and seven samples are selected. These become DCT coefficient information Pi. The above Pg, Ph, and Pi are encoded and multiplexed by the encoder 35 and sent to the receiving side as a digital sequence. The bit allocation per frame for each parameter at this time is as shown in Table 2. However, the DCT coefficient information Pi is composed of the maximum value among the seven samples, seven samples normalized by the maximum value, and position information (grid number).

[Table 2] Since one frame is 20 msec, the encoding speed is 9.2 Kbps, making it possible to achieve a low bit rate.

In FIG. 1B, the received digital sequence is separated into DCT coefficient information Pj, pitch information Pk, and spectrum envelope information Pm by a separation circuit 41, and then DC
A T-coefficient interpolator 42 reproduces DCT coefficient j. Here, as shown in Fig. 2(C), the sent DCT coefficients (7 samples) are rearranged to the original frequency position, and the DCT
Insert 0 into sample points where no coefficients exist, or insert values obtained by interpolation processing. Interpolation method 1
As an example, FIG. 2(D) shows a case where linear interpolation is used. Here, only the components thinned out at equal intervals are subjected to linear interpolation, and 0 is inserted into the other thinned components, resulting in 40 samples. Next, the inverse DCT transform (IDCT) unit 43 transforms into the time domain to reproduce the long-term prediction residual signal k. next,
The long-term prediction synthesizer 44 adds pitch information Pk to the long-term prediction residual signal k and reproduces the short-term prediction residual signal m. Next, the short-term prediction synthesizer 45 adds spectral envelope information Pm to the short-term prediction residual signal m and outputs a reproduced audio signal.

The conversion formulas for DCT and IDCT are as follows, assuming that the input signal is X(n). (1)
In the case of DCT, the required DCT coefficient Xc(k) is, where N is the number of samples per block g(k)=1(k=
0) g(k)=√2(k=1,2...,N-1)(2) I
In the case of DCT, the restored signal X(n) is

0012
]

[Effects of the Invention] As explained in detail above, by implementing the present invention, aliasing distortion does not occur because DCT coefficients are thinned out in the frequency domain, and the quality of reproduced audio is improved compared to conventional methods. . Further, although there is a slight quality deterioration, it has extremely large effects such as being able to lower the encoding speed to 9.2 Kbps.

[Brief explanation of the drawing]

[Fig. 1] Block diagram showing an embodiment of the present invention

[Fig. 2] Explanatory diagram of the thinning method of the present invention

[Figure 3] Block diagram of a conventional audio code/decoder

[Figure 4]
Diagram explaining the occurrence of aliasing distortion

[Explanation of symbols]

11 Short-term prediction analyzer 12 Long-term prediction analyzer 13 LPF 14 Grid selector 15 Encoder 21 Separation circuit 22 Long-term prediction residual regenerator 23 Long-term prediction synthesizer 24 Short-term prediction synthesizer 31 Short-term prediction analyzer 32 Long-term prediction analyzer 33 Discrete cosine transform (DCT) unit 34 Decimator 35 Encoder 41 Separation circuit 42 DCT coefficient interpolator 43 Inverse discrete cosine transform (IDCT) unit 44 Long-term prediction combiner 45 Short-term prediction combiner

Claims (3)

[Claims] 1. Extracting spectral envelope information from an input audio signal by short-term predictive analysis, generating a short-term predictive residual signal by removing the spectral envelope information, and extracting pitch information from the short-term predictive residual signal by long-term predictive analysis. A long-term prediction residual signal is generated by extracting the pitch information and removing the pitch information, and it is converted into the frequency domain by discrete cosine transform to output DCT coefficients which are frequency components.
DCT is performed by removing 2/3 of the T coefficient from the high range and thinning out the remaining 1/3 of the DCT coefficient at equal intervals.
Coefficient information is output, and the DCT coefficient information, the pitch information, and the spectral envelope information are encoded in the form of a digital string signal, multiplexed, and sent to a transmission path, and the digital signal received via the transmission path is The column signal is separated, the DCT coefficient information, the pitch information, and the spectral envelope information are extracted, and the DCT coefficient information is reproduced and D
After rearranging the CT coefficients to their original frequencies, all frequency components are reproduced by inserting all 0s in place of the thinned out components or by inserting values obtained by interpolation, and then by inverse cosine transformation. The long-term prediction residual signal is converted into the time domain and reproduced, the pitch information is added to the long-term prediction residual signal by long-term prediction synthesis to reproduce the short-term prediction residual signal, and then the short-term prediction residual signal is reproduced by short-term prediction synthesis. An audio encoding/decoding method in which an audio signal is decoded and reproduced by adding the spectral envelope information to a prediction residual signal. 2. A short-term prediction analyzer that divides and outputs an input audio signal into spectral envelope information and a short-term prediction residual signal obtained by removing the spectral envelope component from the audio signal, and divides and outputs the short-term prediction residual signal into pitch information and a pitch component. a long-term prediction analyzer that divides and outputs the long-term prediction residual signal from which the long-term prediction residual signal is removed; a discrete cosine transformer that converts the long-term prediction residual signal into the frequency domain and outputs DCT coefficients; Eliminate 2/3 of the component and remove the remaining 1/3
a decimator that decimates DCT coefficients at equal intervals to extract DCT coefficient information; and a code that encodes the spectral envelope information, pitch information, and DCT coefficient information in the form of a digital sequence signal, multiplexes the signal, and sends the multiplexed signal to a transmission path. A speech encoding device equipped with a converter. 3. Receive a multiplexed signal encoded in the form of a digital sequence signal including DCT coefficient information, pitch information, and spectral envelope information, separate the digital sequence signal, and extract the DCT coefficient information, a separator for extracting the pitch information and the spectral envelope information; and the DCT
DCT coefficient interpolation that reproduces the DCT coefficients of all frequency components by rearranging the coefficient information to the original frequency and inserting 0 in place of the thinned out frequency components or by inserting values obtained by interpolation. an inverse cosine transformer that performs an inverse cosine transform on the DCT coefficients and converts them into the time domain to reproduce a long-term prediction residual signal; An audio decoding device comprising: a long-term prediction synthesizer for reproducing; and a short-term prediction synthesizer for decoding and reproducing an audio signal by adding the spectral envelope information to the short-term prediction residual signal.