JP2007072264A - Speech quantization method, speech quantization device, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reproduce high quality broadband speech only by transmitting a few bits of additional information using a telephone band communication channel. <P>SOLUTION: An broadband LSP parameter as envelope information of speech is quantized by: using a certain length of a speech signal as a frame length and analyzing the broadband speech signal and the telephone band speech signal corresponding thereto and limited to a band narrower than that of the broadband signal, respectively; calculating the LSP parameter of the telephone band speech signal and the LSP parameter of the broadband speech signal; selecting some candidates from among a broadband LSP parameter code book using the LSP parameter of the telephone band speech signal, to the broadband LSP parameter code book for quantizing the LSP parameter of the broadband speech signal; and selecting a candidate nearest to the LSP parameter of the input band speech signal from among these candidates selected from the broadband LSP parameter code book. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an LSP quantization method, apparatus, and program for expressing a speech spectrum envelope required when speech is encoded.

Web conferences and multipoint communication conferences can be easily performed in a computer using a personal computer. However, most of these communications use conventional telephone band voice. Several wideband coding methods have already been standardized and can be used. However, although it is used for some services, in general, there are few examples where wideband voice is used. The reason for this is the problem of compatible connection with existing equipment.
That is, in order to communicate with a conventional terminal, it is necessary to use a telephone band voice signal defined by a standard such as G.711. Even in multipoint communication, the wideband coding method cannot be used if there is only one existing terminal that supports only telephone band audio signals. At present, there are many existing telephone band terminals, and it is impossible to replace all terminals with terminals compatible with wideband audio signals. For this reason, it is desirable to be able to connect to a conventional terminal, and in the case of terminals compatible with a wide band, it is desirable to be able to obtain wide band audio.

Non-Patent Document 1 and Non-Patent Document 2 propose a band expansion method for expanding a telephone band audio signal (3.4 kHz band) to a wideband audio signal (7 kHz band). In these methods, a wideband voice signal is generated only from the transmitted telephone band voice signal. At the time of band expansion, it is very important to obtain the spectrum envelope of a wideband speech signal from the speech spectrum envelope (LSP parameter) of the telephone band. In these methods, mapping is used as a method for converting the spectral envelope of a voice signal from a telephone band to a wide band. That is, a codebook for a telephone band envelope spectrum and a codebook for a wideband envelope vector learned for the same voice section are prepared in advance, and only telephone band voice information is sent from the transmission side. The spectral envelope information of the voice obtained from the transmitted telephone band voice is extracted, and the closest code is selected from the envelope spectral codebook for the telephone band. Since this code is the envelope information of the telephone band voice signal, the code of the wideband codebook associated with this code is used as the spectrum envelope of the wideband voice signal.
Yuki Yoshida, Akinobu Abe, “Generation of Wideband Speech from Narrowband Speech Using Codebook Mapping,” IEICE Theory, (D-2), Vol. J78-D-2, No. 3, pp.391-399, March. 1995. Nakahisa Yoshihisa, Tsushima Mineo, Norimatsu Takeshi, “Bandwidth-limited Voice Broadband Using Spectral Linear Mapping,” IEICE Theory, (D-2), Vol. J83-D-2, No. 11, pp. 2246− 2254, Nov. 2000.

According to the band expansion methods proposed in Non-Patent Document 1 and Non-Patent Document 2 described above, if the spectrum envelope of the telephone band is determined, the broadband spectrum envelope is also uniquely determined. However, even if the shape of the spectrum envelope in the telephone band is the same, a band having a higher frequency is not necessarily determined uniquely, and this method cannot achieve high quality.
The object of the present invention is that it is possible to connect to terminals that respond only to the telephone band, and it is obvious that it is possible to communicate with a broadband audio signal between terminals compatible with a wideband, especially a broadband audio signal. It is an object of the present invention to propose a speech quantization method and apparatus that can be decoded with high quality.

  In a system in which telephone band voice signals are exchanged between the transmitting side and the receiving side, the length on the transmitting side is set as the frame length, and the wideband voice signal and the corresponding telephone band voice signal are analyzed to determine the LSP of the telephone band voice signal. Parameters and the LSP parameters of the wideband speech signal are calculated, and a wideband LSP parameter codebook for quantizing the LSP parameters of the wideband speech signal is prepared in advance. Candidates (vectors) from the parameter codebook are narrowed down to several candidates by calculating the LSP parameters of the telephone band voice signal and the distortion of each candidate. Among the candidates selected from these wideband LSP parameter codebooks The candidate closest to the LSP parameter of the input wideband audio signal is selected from Quantized to issue.

  Since the LSP parameter of the telephone band voice signal can be obtained also on the receiving side, the wideband LSP parameter codebook and distortion can be calculated. When selecting an LSP parameter that is close to the input broadband speech LSP parameter, the LSP parameter of the telephone bandwidth speech signal is selected in advance from the wideband LSP parameters stored in the broadband LSP parameter codebook in ascending order of distortion. Since it suffices to select and quantize from a small number of wideband LSP parameters, the bits required for quantization can be reduced.

  According to the present invention, since the LSP parameter of the wideband audio signal is efficiently expressed based on the LSP parameter of the telephone band audio signal, the LSP parameter of the wideband audio signal is quantized with a smaller number of bits than the quantization as it is. Can do. Furthermore, even if the number of bits is small, additional bit information is transmitted as wideband information, so that the wideband information can be decoded by the additional bit information on the receiving side. As a result, it is possible to improve the quality of the reproduced voice as compared with the conventional method of decoding the wideband voice signal relying only on the telephone band voice signal.

The speech quantization apparatus according to the present invention can be configured by hardware. However, in reality, an embodiment in which the speech quantization program according to the present invention is installed in a computer and the computer functions as the speech quantization apparatus is the best. is there.
In order for a computer to function as a speech quantization device, the speech quantization program according to the present invention is installed in the computer, and the speech quantization program is decrypted by the CPU provided in the computer. A wideband speech analyzer that calculates the LSP parameters of a wideband speech signal using the length as the frame length, and the LSP parameters of the telephone bandwidth speech signal that is limited to a narrower bandwidth than the bandwidth of the wideband speech signal using the speech signal as a frame length. A telephone band speech analysis unit to be calculated, a wideband LSP parameter codebook for quantizing the LSP parameters of the wideband speech signal, and vectors that are candidates from the wideband LSP parameter codebook using the LSP parameters of the telephone bandwidth speech signal Search unit for selecting several vectors candidates for vectors selected by this search unit Identify the closest to the vector to LSP parameters of a wideband speech signal from within this vector based on the LSP parameters of a wideband speech signal to construct a wideband LSP quantizer for quantizing function as sound quantizer.

Here, an embodiment of a speech quantization apparatus according to the present invention and a decoding apparatus that decodes a code quantized by the speech quantization apparatus will be described.
FIG. 1 shows a speech quantization apparatus 100, and FIG. Both the speech quantization apparatus 100 and the decoding apparatus 200 are composed of a telephone band processing unit and a broadband processing unit. The voice quantization apparatus 100 limits the bandwidth of the wideband voice signal to a bandwidth of, for example, 3.4 kHz by the low-pass filter 11, down-samples it by the down sampler 12, and then encodes it by the telephone band quantization unit 13.

The wideband quantization unit 15 encodes information necessary for generating high-frequency speech by using the wideband speech signal and the decoded speech signal obtained by decoding the code encoded by the telephone band quantization unit 13.
The telephone band code encoded by the telephone band quantization unit 13 and the extension code encoded by the wideband quantization unit 15 are packetized by the packetization unit 16 and transmitted as transmission data.
The decoding apparatus 200 separates the packet signal sent thereto into a telephone band code and an extension code by the packet separation unit 17. The telephone band code is decoded by the telephone band decoding unit 18 to generate decoded speech. The wideband decoding unit 20 generates a wide area voice signal by using the decoded voice and the extension code of the telephone band decoding unit 18. Only the high frequency band is extracted from the generated high frequency band audio signal by the high-pass filter 19B and added to the telephone band audio signal upsampled by the upsampler 19A to obtain a wideband audio signal.

The configuration of the wideband quantization unit 15 used in the speech quantization apparatus 100 is shown in FIG. The flow of processing is as follows. The telephone band voice signal decoded by the telephone band decoding unit 21 is LPC analyzed by the telephone band voice analysis unit 23 to obtain LSP parameters. Similarly, the wideband audio signal is subjected to LPC analysis by the wideband audio analyzer 22 to obtain LSP parameters of the wideband audio signal. The broadband LSP quantization unit 24 quantizes the wideband LSP parameter using the LSP parameter of the telephone band voice signal and the LSP parameter of the wideband voice signal, and generates an index index as extended information.
The configuration of the wideband decoding unit 20 of the decoding unit 200 is shown in FIG. The flow of processing is as follows.

  The telephone band voice signal is subjected to LPC analysis by the telephone band voice analyzer 32 to obtain LSP parameters. Also, the broadband LSP parameter is obtained by the broadband LSP generation unit 33 using the transmitted index index and the LSP parameter of the telephone band voice signal. The wideband LSP parameters are converted into LPC coefficients and set in the inverse filter 35 and the synthesis filter 37. The telephone band voice signal is upsampled by the upsampler 34 and then passed through the inverse filter 35 to generate a telephone band residual signal. Next, the high frequency residual signal generation unit 36 generates a high frequency component residual signal from the telephone band residual signal by non-linear processing, and the synthesis filter 37 generates high frequency audio. This high frequency audio signal is extracted by the high pass filter 38 and added to the telephone band signal decoded by the telephone band decoding unit 18 shown in FIG. 2 to generate a wide band signal.

  Details of the wideband LSP quantization unit 24 shown in FIG. 3 are shown in FIG. As shown in FIG. 5, the wideband LSP quantization unit 24 uses the LSP parameter Y obtained by analyzing the telephone band voice signal and the LSP parameter X obtained by analyzing the wideband voice signal. The most suitable code is selected from the wideband LSP parameter codebook 41. The characteristics of the low-order part of each candidate vector Zi (M candidate) of the wideband LSP parameter codebook 41 correspond to the LSP parameter Y of the telephone band voice signal. For example, the LSP parameter Y of the telephone band voice signal is an eighth-order vector, and the LSP parameter X of the wideband voice signal is a sixteenth-order vector.

  First, in order to select the candidate vector Zi of the wideband LSP parameter codebook 41 that is close to the shape of the LSP parameter Y of the telephone band voice signal, the distortions of both are calculated for all candidates (M). Next, the top N candidates with the smallest distortion are selected from all candidates. Since a code book corresponding to the wideband LSP parameter code book 41 is also provided on the receiving side, this processing can be similarly performed on the receiving side. Distortion between the top N candidate vectors and the LSP parameter of the actually input wideband speech signal is calculated, and information on the candidate with the smallest distortion is transmitted as an index. For example, by pre-selecting the candidate M = 128 (7 bits) of the broadband LSP parameter codebook 41 with the LSP parameter of the telephone band signal, it is not necessary to transmit all 7 bits as an index. It is only necessary to transmit an index index of N = 8 to 16 (3 to 4 bits).

  That is, on the receiving side, the same wideband as the candidate selected on the transmitting side from the wideband LSP parameter codebook provided in the wideband LSP generating unit 33 (FIG. 4) using the transmitted index index and the decoded telephone band signal. LSP parameters can be extracted. The extracted wideband LSP parameter coincides with the LSP parameter determined to be the LSP parameter closest to the LSP parameter of the wideband audio signal actually input on the transmission side. Therefore, it is possible to always decode a high-quality broadband signal by decoding the high-frequency component using the broadband LSP parameter. As a result, according to the present invention, a high-quality broadband signal can be reproduced simply by adding and transmitting an index index of a small number of bits.

  As a distortion calculation method for each candidate vector Zi (j) of the wideband LSP parameter codebook 41 and the LSP parameter vector Y (j) of the telephone band signal, as shown in FIG. 6, each candidate vector Zi (j) and telephone band signal The distortions between the LSP parameter vectors Y (j) are calculated in order from the lower order vectors, the sum is obtained, and N candidate vectors are extracted from the smaller sum. This search method is called an order search method. This ordered search method is expressed by a mathematical formula:

となる。

It becomes.

  As a method of calculating the distortion of each candidate vector Zi (j) of the wideband LSP parameter codebook 41 and the LSP parameter vector Y (j) of the telephone band signal, as shown in FIG. On the other hand, the distortion is obtained for all combinations of the elements of the candidate vector Zi (j) of the wideband LSP parameter codebook 41, and the combination of the elements of the candidate vector that gives the smallest distortion to the LSP parameter element of the telephone band voice signal is obtained. As a pair, the sum of the distortion of the pair that gives the smallest distortion for all elements of the telephone band signal LSP parameter is obtained, and the distortion of the candidate vector is calculated. Select. This selection method is called a minimum strain search method. This minimum strain search method is expressed by a mathematical formula:

となる。

It becomes.

  The wideband LSP parameter codebook 41 is not a one-stage codebook but a multistage codebook. It is assumed that information after the second stage is transmitted. For example, in the case of a two-stage configuration, it is expressed by the following formula.

The wideband LSP parameter codebook 41 is not obtained only for the corresponding frame, but is quantized using weighting for the previous frame. For example, the LSP parameter Ωn of the nth frame is quantized by the following equation using primary interframe prediction.
Ωn ＝ G0 ・ Cn ＋ G1 ・ Cn−1
Gi is a fixed prediction coefficient, G0 is a value for the current frame, G1 is a value for the previous frame, and Cn is an output vector of the LSP codebook.
The speech quantization apparatus according to the present invention described above can be configured by hardware. However, in order to achieve a more simple implementation, the speech quantization program according to the present invention is installed in a computer, and the speech quantization program is installed in the computer. An embodiment in which a computer is made to function as a speech quantizer by causing the CPU provided in the above to decode and execute is the best.

  The voice quantization program according to the present invention is written in a computer-readable program language and recorded on a recording medium such as a magnetic disk or CD-ROM that can be read by the computer. To install in a computer, it is installed from these recording media or through a communication line.

  It is used in the field of telephone conference equipment or the field of IP telephone using the Internet.

The block diagram for demonstrating the whole audio | voice quantization apparatus by this invention. The block diagram for demonstrating the decoding part which decodes the code | cord | champ quantized with the audio | voice quantization apparatus by this invention. The block diagram for demonstrating the outline | summary of the wideband quantization part used for the audio | voice quantization apparatus by this invention. The block diagram for demonstrating the structure of the wideband decoding part used with the decoding part shown in FIG. The block diagram for demonstrating the search part used for the audio | voice quantization apparatus by this invention. The figure for demonstrating an example of the search method used with the search part shown in FIG. The figure for demonstrating the other example of the search method used with the search part shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Speech quantization apparatus 22 Wideband speech analysis part 11 Low pass filter 23 Telephone band speech analysis part 12 Downsampler 24 Wideband LSP quantization part 13 Telephone band quantization part 31 Telephone band decoding part 14 Delay device 32 Telephone band speech analysis part 15 Wideband Quantization unit 33 Wideband LSP generation unit 16 Packetization unit 34 Upsampler 17 Packet separation unit 35 Inverse filter 18 Telephone band decoding unit 36 Highband residual signal generation unit 19A Upsampler 37 Synthesis filter 19B Highpass filter 38 Highpass filter 20 Wideband decoding Unit 41 Wideband LSP parameter codebook 21 Telephone band decoding unit 42 LSP search unit

Claims (7)

  An audio signal having a certain length as a frame length is analyzed for a wideband audio signal and a corresponding telephone band audio signal that is limited to a band narrower than the band of the wideband audio signal. A wideband LSP parameter codebook for calculating the LSP parameters of a wideband speech signal and quantizing the LSP parameters of the wideband speech signal, using the LSP parameters of the telephone band speech signal as candidates from the wideband LSP parameter codebook And select the candidate closest to the LSP parameter of the input wideband speech signal from the candidates selected from these wideband LSP parameter codebooks, thereby quantizing the LSP parameter, which is the envelope information of speech. A speech quantization method characterized by comprising: The speech quantization method according to claim 1, wherein
When selecting a candidate vector using the telephone band voice LSP parameter vector from the wideband LSP parameter codebook, calculate the mutual distortion of each vector in order from the lowest order, and calculate the sum of each distortion. A speech quantization method, wherein a predetermined number N is selected in order from the smallest. The speech quantization method according to claim 1, wherein
When the search unit selects a candidate vector using the LSP parameter vector of the telephone band voice signal from the wideband LSP parameter codebook, for each element of the LSP parameter of the telephone band voice signal, the element of the wideband LSP parameter codebook Obtain distortion for all combinations, pair the elements of the candidate vector that give the smallest distortion with respect to the LSP parameter element of the telephone band voice signal, and the smallest for all elements of the LSP parameter of the telephone band voice signal The sum of the pair of strains that give the strain is obtained, the strain is calculated with the candidate vectors, and each candidate vector is calculated in the same manner, and the predetermined number N is selected in order from the smallest sum of the strains. Voice quantization method. A wideband speech analysis unit that calculates a LSP parameter of a wideband speech signal with a certain length of the speech signal as a frame length;
A telephone band voice analysis unit for calculating an LSP parameter of a telephone band voice signal limited to a band narrower than a band of the wideband voice signal with a certain length as a frame length of the voice signal;
A wideband LSP parameter codebook for quantizing the LSP parameters of the wideband speech signal;
A search unit that selects several candidate vectors from a wideband LSP parameter codebook using the LSP parameters of the telephone band voice signal;
A wideband quantization unit that identifies a vector that is closest to the LSP parameter of the wideband speech from among the vector candidates selected by the search unit, and that quantizes the LSP parameter of the wideband speech signal based on this vector; ,
A speech quantization apparatus comprising: The speech quantization apparatus according to claim 4, wherein
When the search unit selects candidate vectors from the broadband LSP parameter codebook using LSP parameter vectors of the telephone band voice signal, it calculates the distortion of each vector in order from the lowest order, and obtains the sum of the distortions. A speech quantizing device that selects a predetermined number N in order from the smallest value of the sum of distortions. The speech quantization apparatus according to claim 4, wherein
When the search unit selects a candidate vector from the wideband LSP parameter codebook using the LSP parameter vector of the telephone band voice signal, the search unit selects a candidate vector of the wideband LSP parameter codebook for each element of the LSP parameter of the telephone band voice signal. Find the distortion for all combinations of elements, pair the candidate vector elements that give the least distortion to the LSP parameter elements of the telephone band voice signal, and the smallest for all the elements of the LSP parameter of the telephone band voice signal Calculate the sum of the pair of strains giving the strain, calculate the strain with the candidate vector, calculate each candidate vector in the same way, and select from the smallest value of the sum of strains to the predetermined N in order. A speech quantizer characterized by the above.   7. A program written in a program language that can be read by a computer, and causing the computer to function as the apparatus according to claim 4.

Images