JP2007279754A - Speech encoding device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To encode a speech without causing any local allophone. <P>SOLUTION: A speech encoding device comprises a weighting LPC synthesis unit 306 which obtains a synthesized sound by filtering an adaptive sound source and a probabilistic sound by using an LPC coefficient found from an input speech 301, a gain arithmetic unit 308 which finds gains of the adaptive sound source and probabilistic sound source, a parameter encoding unit 309 which performs vector quantization of the adaptive sound source, probabilistic sound source, and gains found by using encoding distortion between the input speech 301 and synthesized speech, a pitch analyzing unit 310 which finds a correlation value by analyzing pitches of a plurality of subframes before an adaptive code book is searched for a starting subframe of a frame, and calculates the representative pitch closest to a pitch period by using the correlation value, and a search range setting unit 311 which sets a search range of lags of a plurality of subframes based upon the correlation value and representative pitch period, thereby making an adaptive code book search within the search range of lags. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a speech coding apparatus used in a digital communication system.

  In the field of digital mobile communications such as cellular phones, low bit rate speech compression and coding methods are required to cope with the increase in subscribers, and research and development are progressing in each research institution.

  In Japan, the VSELP encoding method developed by Motorola with a bit rate of 11.2 kbps was adopted as the standard encoding method for digital mobile phones, and digital mobile phones equipped with this method were released in Japan in the fall of 1994. Has been.

  In addition, an encoding method called PSI-CELP having a bit rate of 5.6 kbps developed by NTT Mobile Communication Network Corporation has been commercialized. Each of these systems is an improvement on the system called CELP (Code Exited Linear Prediction: M.R.Schroeder "High Quality Speech at Low Bit Rates" Proc.ICASSP '85 pp.937-940).

  This CELP method separates speech into sound source information and vocal tract information, encodes the sound source information by an index of a plurality of sound source samples stored in the codebook, and uses LPC (Linear Prediction Coefficient) for the vocal tract information. It is characterized by adopting a method of encoding and comparing with input speech in consideration of vocal tract information at the time of sound source information encoding (ABS). is there.

  In this CELP method, first, autocorrelation analysis and LPC analysis are performed on input speech data (input speech) to obtain LPC coefficients, and the obtained LPC coefficients are encoded to obtain LPC codes. . Further, the obtained LPC code is decoded to obtain a decoded LPC coefficient. On the other hand, the input voice is perceptually weighted using a perceptual weighting filter using LPC coefficients.

  Obtained for each code vector of the sound source samples (referred to as the adaptive code vector (or adaptive sound source) and the stochastic code vector (or probabilistic sound source)) stored in the adaptive code book and the stochastic code book, respectively. Filtering is performed using the decoded LPC coefficients to obtain two synthesized sounds.

  Then, the relationship between the obtained two synthesized sounds and the auditory weighted input sound is analyzed, the optimum value (optimum gain) of the two synthesized sounds is obtained, and the synthesized sound is power-adjusted with the obtained optimum gain. Then, the synthesized sounds are added to obtain a synthesized synthesized sound. Thereafter, a coding distortion between the obtained synthetic speech and the input speech is obtained. In this way, the coding distortion between the total synthesized sound and the input speech is obtained for all sound source samples, and the index of the sound source sample when the coding distortion is the smallest is obtained.

  The gain and the index of the sound source sample thus obtained are encoded, and the encoded gain and the sound source sample are sent to the transmission path together with the LPC code. In addition, an actual sound source signal is created from two sound sources corresponding to the gain code and the index of the sound source sample, stored in the adaptive codebook, and at the same time, the old sound source sample is discarded.

  In general, the sound source search for the adaptive codebook and the stochastic codebook is performed in a section (called a subframe) obtained by further dividing the analysis section.

  The gain coding (gain quantization) is performed by vector quantization (VQ) that evaluates gain quantization distortion using two synthesized sounds corresponding to the index of the sound source sample.

  In this algorithm, a vector codebook in which a plurality of representative samples (code vectors) of parameter vectors are stored in advance is created. Next, for the perceptually weighted input speech and the perceptually weighted LPC composite of the adaptive sound source and the stochastic sound source, the encoding distortion is calculated by the following equation 1 using the gain code vector stored in the vector codebook. .

式１
ここで、
Ｅ_n：ｎ番のゲインコードベクトルを用いたときの符号化歪み
Ｘ_i：聴感重み付け音声
Ａ_i：聴感重み付けＬＰＣ合成済み適応音源
Ｓ_i：聴感重み付けＬＰＣ合成済み確率的音源
ｇ_n：コードベクトルの要素（適応音源側のゲイン）
ｈ_n：コードベクトルの要素（確率的音源側のゲイン）
ｎ ：コードベクトルの番号
ｉ ：音源データのインデクス
Ｉ ：サブフレーム長（入力音声の符号化単位）

Formula 1
here,
E _n : Coding distortion when using the nth gain code vector X _i : Auditory weighted speech A _i : Auditory weighted LPC synthesized adaptive sound source S _i : Auditory weighted LPC synthesized stochastic sound source g _n : Code vector Element (Gain on the adaptive sound source side)
h _n : Code vector element (stochastic source side gain)
n: code vector number i: index of sound source data I: subframe length (coding unit of input speech)

Then, by comparing the distortion E _n when using each code vector by controlling the vector codebook, a number of the most strained small code vector and the code vector. Also, the code vector number with the smallest distortion among all the code vectors stored in the vector codebook is obtained, and this is used as the code of the vector.

  Although the equation 1 seems to require a lot of calculation for each n at a glance, it is only necessary to calculate the sum of products for i in advance, so that n can be searched with a small amount of calculation. .

  On the other hand, the speech decoding apparatus (decoder) obtains a code vector by decoding the encoded data by obtaining a code vector based on the code of the transmitted vector.

  Further, further improvements have been made on the basis of the above algorithm. For example, using the fact that the auditory characteristic of human sound pressure is logarithm, the power is logarithmized and quantized, and two gains normalized by the power are VQed. This method is used in the standard system of the Japanese PDC half rate codec. In addition, there is a method (predictive coding) in which the gain parameter is used for interframe correlation. This method is described in ITU-T International Standard G.264. 729. However, even with these improvements, sufficient performance cannot be obtained.

  So far, gain information coding methods using human auditory characteristics and inter-frame correlation have been developed, and gain information can be coded with some efficiency. In particular, although the performance is greatly improved by predictive quantization, in the conventional method, predictive quantization is performed using the value of the previous subframe as it is as the state value. However, some of the values stored as states take extremely large (small) values. If these values are used for the next subframe, the quantization of the next subframe does not work, and the local May be an unusual noise.

  The present invention has been made in view of this point, and an object of the present invention is to provide a CELP speech coding apparatus that can perform speech coding without causing local abnormal noise.

  The present invention is a CELP speech coding apparatus that performs coding by decomposing one frame into a plurality of subframes, and for adaptive and stochastic sound sources stored in the adaptive code book and stochastic code book. Then, LPC synthesis means for obtaining a synthesized sound by filtering using an LPC coefficient obtained from the input voice, gain calculating means for obtaining gains of the adaptive sound source and the stochastic sound source, the input sound and the synthesized sound Before performing the adaptive codebook search of the first subframe of the frame, and the adaptive coding and stochastic excitation obtained using the coding distortion between and the parameter coding means for vector quantization of the gain , Calculating a correlation value by performing a pitch analysis of the plurality of subframes, and calculating a representative pitch value that is the closest to the pitch period using the correlation value. And a search range setting means for setting a search range of lags of the plurality of subframes based on the correlation value and the representative pitch period, and the adaptive codebook search is performed with respect to the search range of the lags. The configuration that is performed is adopted.

  According to the present invention, since the vicinity of the provisional pitch of the second subframe can be searched when searching for the second subframe, the first and second subframes can be obtained even in a non-stationary frame such as when speech starts from the second half of the frame. In this case, an appropriate lag search can be performed, and speech coding can be performed without causing local abnormal noise.

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

(Embodiment 1)
FIG. 1 is a block diagram showing a configuration of a wireless communication apparatus provided with a speech encoding apparatus according to Embodiments 1 and 2 of the present invention.

  In this wireless communication device, the voice is converted into an electrical analog signal by the voice input device 11 such as a microphone on the transmission side, and is output to the A / D converter 12. The analog audio signal is converted into a digital audio signal by the A / D converter 12 and output to the audio encoding unit 13. The speech encoding unit 13 performs speech encoding processing on the digital speech signal and outputs the encoded information to the modem unit 14. The modem unit 14 digitally modulates the encoded audio signal and sends it to the wireless transmission unit 15. The wireless transmission unit 15 performs predetermined wireless transmission processing on the modulated signal. This signal is transmitted via the antenna 16. The processor 21 performs processing using data stored in the RAM 22 and ROM 23 as appropriate.

  On the other hand, on the reception side of the wireless communication apparatus, the reception signal received by the antenna 16 is subjected to predetermined wireless reception processing by the wireless reception unit 17 and is sent to the modulation / demodulation unit 14. The modem unit 14 performs demodulation processing on the received signal and outputs the demodulated signal to the speech decoding unit 18. The audio decoding unit 18 performs a decoding process on the demodulated signal to obtain a digital decoded audio signal, and outputs the digital decoded audio signal to the D / A converter 19. The D / A converter 19 converts the digital decoded audio signal output from the audio decoding unit 18 into an analog decoded audio signal and outputs the analog decoded audio signal to an audio output device 20 such as a speaker. Finally, the audio output device 20 converts the electrical analog decoded audio signal into decoded audio and outputs it.

  Here, the speech encoding unit 13 and the speech decoding unit 18 are operated by a processor 21 such as a DSP using a code book stored in the RAM 22 and the ROM 23. These operation programs are stored in the ROM 23.

  FIG. 2 is a block diagram showing the configuration of the CELP speech coding apparatus according to Embodiment 1 of the present invention. This speech encoding apparatus is included in speech encoding unit 13 shown in FIG. The adaptive codebook 103 shown in FIG. 2 is stored in the RAM 22 shown in FIG. 1, and the probabilistic codebook 104 shown in FIG. 2 is stored in the ROM 23 shown in FIG.

  In the speech encoding apparatus shown in FIG. 2, the LPC analysis unit 102 performs autocorrelation analysis and LPC analysis on the input speech data (input speech) 101 to obtain LPC coefficients. Also, the LPC analysis unit 102 encodes the obtained LPC coefficient to obtain an LPC code. Further, the LPC analysis unit 102 decodes the obtained LPC code to obtain decoded LPC coefficients. The input audio data 101 is sent to the perceptual weighting unit 107 where the perceptual weighting is performed using the perceptual weighting filter using the LPC coefficient.

  Next, in the sound source creation unit 105, a sound source sample (adaptive code vector or adaptive sound source) stored in the adaptive codebook 103 and a sound source sample (stochastic code vector or stochastic sound source) stored in the probabilistic codebook 104 Are sent to the audible weight LPC synthesis unit 106. Further, the audible weight LPC synthesis unit 106 filters the two sound sources obtained by the sound source creation unit 105 with the decoded LPC coefficient obtained by the LPC analysis unit 102 to obtain two synthesized sounds.

  Note that the perceptual weight LPC synthesis unit 106 uses a perceptual weighting filter using an LPC coefficient, a high-frequency emphasis filter, and a long-term prediction coefficient (obtained by performing long-term prediction analysis of the input speech) and uses each synthesized sound. Is subjected to auditory weighting LPC synthesis.

  The audible weight LPC synthesis unit 106 outputs two synthesized sounds to the gain calculation unit 108. The gain calculation unit 108 has the configuration shown in FIG. In the gain calculation unit 108, the two synthesized sounds obtained by the perceptual weight LPC synthesis unit 106 and the perceptually weighted input speech are sent to the analysis unit 1081, where the relationship between the two synthesized sounds and the input speech is analyzed. The optimum value (optimum gain) of the two synthesized sounds is obtained. This optimum gain is output to the power adjustment unit 1082.

  The power adjustment unit 1082 adjusts the power of the two synthesized sounds with the obtained optimum gain. The power-adjusted synthesized sound is output to the synthesizing unit 1083, where it is added to become a total synthesized sound. This total synthesized sound is output to the coding distortion calculation unit 1084. The coding distortion calculation unit 1084 obtains coding distortion between the obtained synthetic speech and input speech.

  The encoding distortion calculation unit 1084 controls the sound source creation unit 105 to output all the sound source samples of the adaptive codebook 103 and the stochastic codebook 104, and for all the sound source samples, the total synthesized sound, the input speech, And the index of the sound source sample when the coding distortion is the smallest.

  Next, the analysis unit 1081 sends the index of the sound source sample, two perceptually weighted LPC synthesized sound sources corresponding to the index, and the input speech to the parameter encoding unit 109.

  The parameter encoding unit 109 obtains a gain code by performing gain encoding, and collectively sends the LPC code and the index of the sound source sample to the transmission path. Further, an actual sound source signal is created from two sound sources corresponding to the gain code and the index, and stored in the adaptive codebook 103, and at the same time, the old sound source sample is discarded. In general, the sound source search for the adaptive codebook and the stochastic codebook is performed in a section (called a subframe) obtained by further dividing the analysis section.

  Here, the operation of gain encoding of parameter encoding section 109 of the speech encoding apparatus having the above configuration will be described. FIG. 4 is a block diagram showing the configuration of the parameter encoding unit of the speech encoding apparatus of the present invention.

In FIG. 4, the perceptually weighted input sound (X _i ), perceptual weighted LPC synthesized adaptive sound source (A _i ), and perceptually weighted LPC synthesized stochastic sound source (S _i ) are sent to the parameter calculator 1091. The parameter calculation unit 1091 calculates parameters necessary for encoding distortion calculation. The parameter calculated by the parameter calculation unit 1091 is output to the coding distortion calculation unit 1092 where the coding distortion is calculated. The encoding distortion is output to the comparison unit 1093. The comparison unit 1093 controls the coding distortion calculation unit 1092 and the vector codebook 1094 to obtain the most appropriate code (decoded vector) from the obtained coding distortion, and based on this code, the vector codebook The code vector obtained from 1094 is output to the decoded vector storage unit 1096, and the decoded vector storage unit 1096 is updated.

  The prediction coefficient storage unit 1095 stores prediction coefficients used for predictive coding. Since this prediction coefficient is used for parameter calculation and coding distortion calculation, it is output to the parameter calculation section 1091 and coding distortion calculation section 1092. The decoded vector storage unit 1096 stores a state for predictive encoding. Since this state is used for parameter calculation, it is output to the parameter calculation unit 1091. The vector codebook 1094 stores code vectors.

Next, an algorithm of the gain encoding method according to the present invention will be described.
A vector codebook 1094 in which a plurality of representative samples (code vectors) of quantization target vectors are stored in advance is created. Each vector includes three elements: an AC gain, a value corresponding to a logarithmic value of the SC gain, and an adjustment coefficient of the SC prediction coefficient.

  This adjustment coefficient is a coefficient that adjusts the prediction coefficient in accordance with the state of the previous subframe. Specifically, this adjustment coefficient is set to reduce the influence when the state of the previous subframe is an extremely large value or an extremely small value. This adjustment coefficient can be obtained by a learning algorithm developed by the present inventors using a large number of vector samples. Here, description of this learning algorithm is omitted.

  For example, a large adjustment coefficient is set for a code vector frequently used for voiced sound. That is, when the same waveform is arranged, since the reliability of the state of the previous subframe is high, the adjustment coefficient is increased so that the prediction coefficient of the previous subframe can be used as it is. Thereby, more efficient prediction can be performed.

  On the other hand, an adjustment coefficient is made small for a code vector that is used infrequently and used less frequently. In other words, if the previous waveform is completely different, the reliability of the previous subframe state is low (the adaptive codebook is considered not to function), so the adjustment coefficient is reduced and the prediction coefficient of the previous subframe is reduced. Reduce the impact of As a result, it is possible to prevent the adverse effect of the next prediction and realize good predictive coding.

  In this way, by controlling the prediction coefficient according to each code vector (state), it is possible to further improve the performance of the prediction encoding so far.

  The prediction coefficient storage unit 1095 stores a prediction coefficient for performing predictive coding. This prediction coefficient is an MA (moving average) prediction coefficient, and stores two types of AC and SC corresponding to the prediction order. These prediction coefficient values are generally obtained in advance by learning using a large amount of data. The decoded vector storage unit 1096 stores a value indicating a silent state as an initial value.

Next, the encoding method will be described in detail. First, the perceptual weighting input speech (X _i ), perceptual weighting LPC synthesized adaptive sound source (A _i ), perceptual weighting LPC synthesized stochastic sound source (S _i ) are sent to the parameter calculation unit 1091, and the decoded vector storage unit 1096. The decoding vector (AC, SC, adjustment coefficient) stored in the prediction coefficient storage unit 1095 and the prediction coefficient (AC, SC) stored in the prediction coefficient storage unit 1095 are sent. These parameters are used to calculate parameters necessary for coding distortion calculation.

  The encoding distortion calculation in the encoding distortion calculation unit 1092 is performed according to the following equation 2.

式２
ここで、
Ｇ_an，Ｇ_sn：復号化ゲイン
Ｅ_n：ｎ番のゲインコードベクトルを用いたときの符号化歪み
Ｘ_i：聴感重み付け音声
Ａ_i：聴感重み付けＬＰＣ合成済み適応音源
Ｓ_i：聴感重み付けＬＰＣ合成済み確率的音源
ｎ ：コードベクトルの番号
ｉ ：音源ベクトルのインデクス
Ｉ ：サブフレーム長（入力音声の符号化単位）

Formula 2
here,
G _an, G _sn: decoding gain E _n: coding distortion when using the gain code vector of n-th X _i: perceptual weighting speech A _i: perceptual weighting LPC precomposed adaptive excitation S _i: perceptual weighting LPC precomposed Probabilistic sound source n: Code vector number i: Sound source vector index I: Subframe length (coding unit of input speech)

In this case, in order to reduce the amount of calculation, the parameter calculation unit 1091 calculates a portion that does not depend on the code vector number. What is calculated is the correlation and power between the prediction vector and the three synthesized sounds (X _i , A _i , S _i ). This calculation is performed according to the following equation 3.

式３
Ｄ_xx，Ｄ_xa，Ｄ_xs，Ｄ_aa，Ｄ_as，Ｄ_ss：合成音間の相関値、パワ
Ｘ_i：聴感重み付け音声
Ａ_i：聴感重み付けＬＰＣ合成済み適応音源
Ｓ_i：聴感重み付けＬＰＣ合成済み確率的音源
ｎ ：コードベクトルの番号
ｉ ：音源ベクトルのインデクス
Ｉ ：サブフレーム長（入力音声の符号化単位）

Formula 3
D _xx , D _xa , D _xs , D _aa , D _as , D _ss : correlation value between synthesized sounds, power X _i : perceptually weighted speech A _i : perceptually weighted LPC synthesized adaptive sound source S _i : perceptually weighted LPC synthesized Probabilistic sound source n: Code vector number i: Sound source vector index I: Subframe length (coding unit of input speech)

  Further, the parameter calculation unit 1091 calculates three prediction values shown in the following equation 4 using the past code vector stored in the decoded vector storage unit 1096 and the prediction coefficient stored in the prediction coefficient storage unit 1095. Keep it.

式４
ここで、
Ｐ_ra：予測値（ＡＣゲイン）
Ｐ_rs：予測値（ＳＣゲイン）
Ｐ_sc：予測値（予測係数）
α_m：予測係数（ＡＣゲイン、固定値）
β_m：予測係数（ＳＣゲイン、固定値）
Ｓ_am：状態（過去のコードベクトルの要素、ＡＣゲイン）
Ｓ_sm：状態（過去のコードベクトルの要素、ＳＣゲイン）
Ｓ_cm：状態（過去のコードベクトルの要素、ＳＣ予測係数調整係数）
ｍ：予測インデクス
Ｍ：予測次数

Formula 4
here,
P _ra : Predicted value (AC gain)
P _rs : Predicted value (SC gain)
P _sc : Prediction value (prediction coefficient)
α _m : Prediction coefficient (AC gain, fixed value)
β _m : Prediction coefficient (SC gain, fixed value)
S _am : State (element of past code vector, AC gain)
S _sm : State (element of past code vector, SC gain)
S _cm : State (element of past code vector, SC prediction coefficient adjustment coefficient)
m: Predicted index M: Predicted order

As can be seen from Equation 4, P _rs and P _sc are multiplied by an adjustment coefficient unlike the conventional case. Therefore, with respect to the prediction value and prediction coefficient of the SC gain, when the value of the state in the previous subframe is extremely large or small, it can be mitigated (the influence is reduced) by the adjustment coefficient. That is, it is possible to adaptively change the SC gain prediction value and the prediction coefficient according to the state.

  Next, the encoding distortion calculation unit 1092 uses the parameters calculated by the parameter calculation unit 1091, the prediction coefficients stored in the prediction coefficient storage unit 1095, and the code vector stored in the vector codebook 1094, and The coding distortion is calculated according to 5.

式５
ここで、
Ｅ_n：ｎ番のゲインコードベクトルを用いたときの符号化歪み
Ｄ_xx，Ｄ_xa，Ｄ_xs，Ｄ_aa，Ｄ_as，Ｄ_ss：合成音間の相関値、パワ
Ｇ_an，Ｇ_sn：復号化ゲイン
Ｐ_ra：予測値（ＡＣゲイン）
Ｐ_rs：予測値（ＳＣゲイン）
Ｐ_ac：予測係数の和（固定値）
Ｐ_sc：予測係数の和（上記式４で算出）
Ｃ_an，Ｃ_sn，Ｃ_cn：コードベクトル、Ｃ_cnは予測係数調整係数であるがここでは使用しない
ｎ：コードベクトルの番号
なお、実際にはＤ_xxはコードベクトルの番号ｎに依存しないので、その加算を省略することができる。

Formula 5
here,
E _n : Coding distortion when using the nth gain code vector D _xx , D _xa , D _xs , D _aa , D _as , D _ss : Correlation value between synthesized sounds, Power G _an , G _sn : Decoding Gain P _ra : Predicted value (AC gain)
P _rs : Predicted value (SC gain)
P _ac : Sum of prediction coefficients (fixed value)
P _sc : sum of prediction coefficients (calculated by the above equation 4)
C _an , C _sn , C _cn : code vector, C _cn is a prediction coefficient adjustment coefficient, but is not used here n: code vector number Note that D _xx does not actually depend on the code vector number n. The addition can be omitted.

  Next, the comparison unit 1093 controls the vector codebook 1094 and the encoding distortion calculation unit 1092, and the encoding calculated by the encoding distortion calculation unit 1092 among the plurality of code vectors stored in the vector codebook 1094 is performed. The code vector number with the smallest distortion is obtained, and this is used as the sign of the gain. Also, the content of decoded vector storage section 1096 is updated using the gain code obtained. The update is performed according to the following formula 6.

式６
ここで、
Ｓ_am，Ｓ_sm，Ｓ_cm：状態ベクトル（ＡＣ、ＳＣ、予測係数調整係数）
ｍ：予測インデクス
Ｍ：予測次数
Ｊ：比較部で求められた符号

Equation 6
here,
S _am , S _sm , S _cm : State vector (AC, SC, prediction coefficient adjustment coefficient)
m: Prediction index M: Prediction order J: Code obtained by the comparison unit

As can be seen from Expression 4 to Expression 6, in this embodiment, the state vector S _cm is stored in the decoded vector storage unit 1096, and the prediction coefficient is adaptively controlled using this prediction coefficient adjustment coefficient. is doing.

  FIG. 5 is a block diagram showing a configuration of the speech decoding apparatus according to the embodiment of the present invention. This speech decoding apparatus is included in the speech decoding unit 18 shown in FIG. The adaptive codebook 202 shown in FIG. 5 is stored in the RAM 22 shown in FIG. 1, and the probabilistic codebook 203 shown in FIG. 5 is stored in the ROM 23 shown in FIG.

  In the speech decoding apparatus shown in FIG. 5, parameter decoding section 201 obtains an encoded speech signal from the transmission path, and excitation samples of each excitation codebook (adaptive codebook 202 and probabilistic codebook 203). , LPC code, and gain code. Then, an LPC coefficient decoded from the LPC code is obtained, and a decoded gain is obtained from the gain code.

  Then, the sound source creation unit 204 obtains a decoded sound source signal by multiplying each sound source sample by the decoded gain and adding the result. At this time, the obtained decoded excitation signal is stored in the adaptive codebook 204 as an excitation sample, and the old excitation sample is discarded at the same time. Then, the LPC synthesis unit 205 obtains a synthesized sound by performing filtering on the decoded sound source signal using the decoded LPC coefficient.

  The two excitation codebooks are the same as those included in the speech encoding apparatus shown in FIG. 2 (reference numerals 103 and 104 in FIG. 2), and sample numbers for extracting excitation samples (adaptive codebooks). Both the code to and the code to the stochastic codebook are supplied from the parameter decoding unit 201.

  As described above, in the speech coding apparatus according to the present embodiment, the prediction coefficient can be controlled according to each code vector, and more efficient prediction adapted to local features of speech and unsteady part This makes it possible to prevent the adverse effects of predictions and to obtain exceptional effects that have not been obtained in the past.

(Embodiment 2)
In the speech coding apparatus, as described above, the gain calculation unit compares the synthesized sound and the input speech for all the sound sources of the adaptive codebook and the stochastic codebook obtained from the sound source creation unit. At this time, normally, two sound sources (adaptive codebook and probabilistic codebook) are searched in an open loop for the convenience of computation. Hereinafter, a description will be given with reference to FIG.

  In this open loop search, first, the sound source creation unit 105 selects sound source candidates one after another only from the adaptive codebook 103, obtains a synthesized sound by causing the perceptual weight LPC synthesis unit 106 to function, and sends it to the gain calculation unit 108. A comparison between the synthesized speech and the input speech is performed to select an optimum code of the adaptive codebook 103.

  Next, the code of the adaptive codebook 103 is fixed, the same sound source is selected from the adaptive codebook 103, and the sound source corresponding to the code of the gain calculation unit 108 is selected one after another from the probabilistic codebook 104. The data is transmitted to the weight LPC synthesis unit 106. The gain calculator 108 compares the sum of both synthesized sounds and the input speech to determine the code of the stochastic codebook 104.

  When this algorithm is used, the coding performance is slightly deteriorated compared to searching all codes of all codebooks, but the calculation amount is greatly reduced. For this reason, this open loop search is generally used.

  Here, a typical algorithm in the conventional open loop sound source search will be described. Here, a sound source search procedure in the case where a single analysis section (frame) is composed of two subframes will be described.

  First, upon receiving an instruction from the gain calculation unit 108, the sound source creation unit 105 extracts a sound source from the adaptive codebook 103 and sends it to the audible weight LPC synthesis unit 106. In gain calculation section 108, the comparison between the synthesized sound source and the input speech of the first subframe is repeated to obtain the optimum code. Here, features of the adaptive codebook are shown. The adaptive codebook is a sound source used for synthesis in the past. And the code | symbol respond | corresponds to a time lag as shown in FIG.

  Next, after the codes of the adaptive codebook 103 are determined, the probabilistic codebook is searched. The sound source creation unit 105 extracts the sound source of the code obtained by the search of the adaptive code book 103 and the sound source of the probabilistic code book 104 specified by the gain calculation unit 108 and sends them to the audible weight LPC synthesis unit 106. Then, gain calculation section 108 calculates coding distortion between the perceptually weighted synthesized sound and perceptually weighted input speech, and the most appropriate (the one with the least square error) code of stochastic sound source 104. Decide. The excitation code search procedure in one analysis section (when the subframe is 2) is shown below.

1) Determine the code of the adaptive codebook of the first subframe 2) Determine the code of the stochastic codebook of the first subframe 3) The gain is encoded by the parameter encoding unit 109, and the first subframe is determined by the decoding gain. And the adaptive codebook 103 is updated.
4) Determine the code of the adaptive codebook of the second subframe 5) Determine the code of the stochastic codebook of the second subframe 6) The gain is encoded by the parameter encoding unit 109, and the second subframe is determined by the decoding gain. And the adaptive codebook 103 is updated.

  The sound source can be efficiently encoded by the above algorithm. Recently, however, efforts have been made to save the number of bits of a sound source in order to further reduce the bit rate. Of particular interest is the fact that there is a large correlation in the lag of the adaptive codebook, so that the code of the first subframe remains unchanged and the search range of the second subframe is close to the lag of the first subframe. The number of bits is reduced (by reducing the number of entries).

  In this algorithm, it is conceivable that local degradation is caused when the voice changes from the middle of the analysis section (frame) or when the states of the two subframes are greatly different.

  In the present embodiment, a search method for calculating a correlation value by performing pitch analysis for both two subframes before encoding and determining a search range of lags of the two subframes based on the obtained correlation value A speech encoding device that achieves the above is provided.

  Specifically, the speech coding apparatus according to the present embodiment is a CELP type coding apparatus that decomposes one frame into a plurality of subframes and codes each of them, and performs adaptive codebook search for the first subframe. Before, a pitch analysis unit that calculates a correlation value by performing a pitch analysis of a plurality of subframes constituting a frame, and the pitch analysis unit calculates a correlation value of a plurality of subframes that constitute a frame, and the correlation The value that is most likely to be the pitch period in each subframe (called the representative pitch) is obtained from the magnitude of the value, and the lag search range for multiple subframes is determined based on the correlation value obtained by the pitch analyzer and the representative pitch And a search range setting unit.

  In this speech encoding apparatus, the search range setting unit uses a representative pitch and a correlation value of a plurality of subframes obtained by the pitch analysis unit to use a temporary pitch that is the center of the search range (referred to as a temporary pitch). In the search range setting unit, the search range of the lag is set in a specified range around the calculated temporary pitch, and the search range is set before and after the temporary pitch when the search interval of the lag is set. At that time, candidates for short lag portions are reduced, a longer lag range is set wider, and lag search is performed within the range set by the search range setting unit during adaptive codebook search.

  Hereinafter, the speech coding apparatus according to the present embodiment will be described in detail with reference to the accompanying drawings. Here, it is assumed that one frame is divided into two subframes. Even in the case of three or more subframes, encoding can be performed in the same procedure.

  In this speech coding apparatus, in the pitch search by the so-called delta lag method, the pitch is obtained for all the divided subframes, the degree of correlation between the pitches is obtained, and the search range is determined according to the correlation result. To decide.

  FIG. 7 is a block diagram showing the configuration of the speech coding apparatus according to Embodiment 2 of the present invention. First, the LPC analysis unit 302 obtains LPC coefficients by performing autocorrelation analysis and LPC analysis on the input voice data (input voice) 301. The LPC analysis unit 302 encodes the obtained LPC coefficient to obtain an LPC code. Further, the LPC analysis unit 302 decodes the obtained LPC code to obtain decoded LPC coefficients.

Next, the pitch analysis unit 310 performs pitch analysis of input speech for two subframes to obtain pitch candidates and parameters. The algorithm for one subframe is shown below. Two correlation coefficients are obtained by the following equation (7). Note In this, C _pp is determined first for P _min, for _{P min + 1, P min +} 2 after, can be efficiently computed by pulling adding the value of the frame end.

式７
ここで、
Ｘ_i，Ｘ_i-P：入力音声
Ｖ_p：自己相関関数
Ｃ_pp：パワ成分
ｉ：入力音声のサンプル番号
Ｌ：サブフレームの長さ
Ｐ：ピッチ
Ｐ_min，Ｐ_max：ピッチの探索を行う最小値と最大値

Equation 7
here,
X _i , X _iP : Input speech V _p : Autocorrelation function C _pp : Power component i: Sample number of input speech L: Length of subframe P: Pitch P _min , P _max : Minimum value for searching for pitch Maximum value

Then, the autocorrelation function and the power component obtained by the above equation 7 are stored in the memory, and the representative pitch P ₁ is obtained by the following procedure. This is a process for obtaining the pitch P that maximizes V _p × V _p / C _pp when V _p is positive. However, since division generally requires a calculation amount, both the numerator and the denominator are stored, and the efficiency is improved by multiplying them.

Here, a search is made for a pitch that minimizes the sum of squares of the difference between the input sound and the input sound from the adaptive sound source that is past the pitch. This process is equivalent to the process for obtaining the pitch P that maximizes V _p × V _p / C _pp . Specific processing is as follows.

1) Initialization _{(P = P min, VV =} C = 0, P 1 = P min)
2) If the _{(V p × V p × C} <VV × C pp) or (V _p <0) if 4). Otherwise go to 3).
_{3) VV = V p × V} p, C = C pp, 4 as P ₁ = P 4) to) and P = P + 1. At this time, if P> P _max , the process ends. Otherwise, go to 2).

The above operation is performed for each of the two subframes to obtain representative pitches P ₁ and P ₂ , autocorrelation coefficients V _1p and V _2p , and power components C _1pp and C _2pp (P _min <p <P _max ).

  Next, the search range setting unit 311 sets the search range of the lag of the adaptive codebook. First, a temporary pitch that is the axis of the search range is obtained. The provisional pitch is determined using the representative pitch and parameters obtained by the pitch analysis unit 310.

The temporary pitches Q ₁ and Q ₂ are obtained by the following procedure. In the following description, a constant Th (specifically, about 6 is appropriate) is used as the range of the lag. Further, the correlation value obtained by Equation 7 is used.

First, a temporary pitch (Q ₂ ) having the largest correlation is found near P ₁ (± Th) with P ₁ fixed.

1) Initialization (p = P ₁ −Th, C _max = 0, Q ₁ = P ₁ , Q ₂ = P ₁ )
2) If (V _1p1 × V _1p1 / C _1p1p1 + V _2p × V _2p / C _2pp <C _max ) or (V _2p <0), go to 4). Otherwise go to 3).
_{3) C max = V 1p1 ×} V 1p1 / C 1p1p1 + V 2p × V 2p / C 2pp, Q 2 = p as 4) to 4) p = p + 1 as a 2) to. However, if p> P ₁ + Th at this time, go to 5).

In this way, the processing of 2) to 4) is performed from P ₁ −Th to P ₁ + Th, and the _maximum correlation C _max and provisional pitch Q ₂ are obtained.

Next, a temporary pitch (Q ₁ ) having the largest correlation is obtained in the vicinity of P ₂ (± Th) with P ₂ fixed. In this case, C _max is not initialized. By obtaining Q ₁ that maximizes the correlation including C _max when Q ₂ is obtained, it is possible to obtain Q ₁ and Q ₂ having the maximum correlation between the first and second subframes. .

5) Initialization (p = P ₂ −Th)
6) If (V _1p × V _1p / C _1pp + V _2p2 × V _2p2 / C _2p2p2 <C _max ) or (V _1p <0) then go to 8). Otherwise go to 7).
_{7) C max = V 1p ×} V 1p / C 1pp + V 2p2 × V 2p2 / C 2p2p2, Q 1 = p, Q 2 = P 2 as 8) to.
8) Go to 6) with p = p + 1. However, if p> P ₂ + Th at this time, go to 9).
9) End.

Thus the to 6) to 8) of the processing performed up to P ₂ -Th~P ₂ + Th, finding the largest one C _max and provisional pitch Q _1, Q ₂ of the correlation. Q ₁ and Q ₂ at this time are provisional pitches of the first subframe and the second subframe.

With the above algorithm, it is possible to select two provisional pitches that have relatively little difference in size (the maximum difference is Th) while simultaneously evaluating the correlation between two subframes. By using this provisional pitch, it is possible to prevent the coding performance from being greatly deteriorated even when the search range is set narrow during the adaptive codebook search of the second subframe. For example, when the sound quality suddenly changes from the second subframe and the correlation of the second subframe is strong, the deterioration of the second subframe is reduced by using Q ₁ reflecting the correlation of the second subframe. Can be avoided.

Furthermore, search range setting section 311 sets the range of searching the adaptive codebook using provisional pitch Q ₁ which was determined (L_ _ST ~L_ _EN) as the following equation 8.

式８
ここで、
Ｌ__ST：探索範囲の始点
Ｌ__EN：探索範囲の終点
Ｌ_min：ラグの最小値（例：２０）
Ｌ_max：ラグの最大値（例：１４３）
Ｔ₁：第１サブフレームの適応符号帳ラグ

Equation 8
here,
L_ _ST: starting point of search range L_ _EN: search range end point L _min: minimum value of lag (e.g. 20)
L _max : Maximum value of lag (example: 143)
T ₁ : Adaptive codebook lag of the first subframe

  In the above setting, it is not necessary to narrow the search range for the first subframe. However, the present inventors have confirmed through experiments that the performance is better when the vicinity of the value based on the pitch of the input speech is set as the search interval. In the present embodiment, the algorithm is searched by narrowing to 26 samples. Is used.

In the second subframe, a search range is set around the lag T ₁ obtained in the first subframe. Therefore, the lag of the adaptive codebook of the second subframe can be encoded with 5 bits with a total of 32 entries. In addition, the present inventors have confirmed by experiments that a better performance can be obtained by setting a small number of candidates with a small lag and a large number of candidates with a large lag. However, as can be seen from the above description, the temporary pitch Q ₂ is not used in the present embodiment.

  Here, the effect in this Embodiment is demonstrated. The temporary pitch of the second subframe also exists near the temporary pitch of the first subframe obtained by the search range setting unit 311 (because it is limited by the constant Th). In addition, since the search is performed by narrowing the search range in the first subframe, the lag obtained as a result of the search does not deviate from the temporary pitch of the first subframe.

  Therefore, when searching for the second subframe, a range close to the provisional pitch of the second subframe can be searched, and an appropriate lag can be searched for in both the first and second subframes.

As an example, let us consider a case where the first subframe is silent and the voice rises from the second subframe. In the conventional method, if the pitch of the second subframe is not included in the search section by narrowing the search range, the sound quality is greatly deteriorated. In the method according to the present embodiment, the correlation of the representative pitch P ₂ is strong in the temporary pitch analysis of the pitch analysis unit. Therefore, the provisional pitch of the first subframe is a value close to P _2. For this reason, in the search by the delta lag, a portion close to the portion where the voice rises can be set as the temporary pitch. That is, when the search of the adaptive codebook of the second subframe will be to explore the value close P _2, to perform an adaptive codebook search of the second subframe by the way without deterioration even if the rising of the voice Derudaragu Can do.

  Next, in the sound source creation unit 305, the sound source sample (adaptive code vector or adaptive sound source) stored in the adaptive codebook 303 and the sound source sample (stochastic code vector or stochastic sound source) stored in the probabilistic codebook 304 are used. Each is taken out and sent to the auditory weight LPC synthesis unit 306. Further, the audible weight LPC synthesis unit 306 performs filtering on the two sound sources obtained by the sound source creation unit 305 with the decoded LPC coefficients obtained by the LPC analysis unit 302 to obtain two synthesized sounds.

  Further, the gain calculation unit 308 analyzes the relationship between the two synthesized sounds obtained by the perceptual weight LPC synthesis unit 306 and the input speech weighted by the perceptual weighting unit 307, and determines the optimum value ( Find the optimal gain. Further, gain calculation section 308 adds the synthesized sounds adjusted for power by the optimum gain to obtain a total synthesized sound. Then, the gain calculation unit 308 calculates the coding distortion of the total synthesized sound and the input speech. In the gain calculation unit 308, many synthesized sounds obtained by causing the sound source creation unit 305 and the auditory weight LPC synthesis unit 306 to function on all the sound source samples of the adaptive codebook 303 and the stochastic codebook 304 Encoding distortion with the input speech is performed, and the index of the sound source sample when the encoding distortion obtained as a result is the smallest is obtained.

  Next, the index of the obtained sound source sample, the two sound sources corresponding to the index, and the input speech are sent to the parameter encoding unit 309. The parameter encoding unit 309 obtains a gain code by performing gain encoding, and sends the gain code together with the LPC code and the index of the sound source sample to the transmission line.

  The parameter encoding unit 309 creates an actual sound source signal from two sound sources corresponding to the gain code and the index of the sound source sample, stores it in the adaptive codebook 303, and simultaneously discards the old sound source sample.

  Note that the perceptual weight LPC synthesis unit 306 uses a perceptual weighting filter that uses an LPC coefficient, a high-frequency emphasis filter, and a long-term prediction coefficient (obtained by performing long-term prediction analysis of input speech).

  The gain calculation unit 308 performs comparison between the input speech for all sound sources of the adaptive codebook 303 and the probabilistic codebook 304 obtained from the sound source creation unit 305. (Adaptive codebook 303 and probabilistic codebook 304) are searched by open loop as described above.

  As described above, the pitch search method in the present embodiment calculates the correlation value by performing the pitch analysis of the plurality of subframes constituting the frame before the adaptive codebook search of the first subframe. The correlation values of all the subframes can be grasped simultaneously.

  Then, the correlation value of each subframe is calculated, and a value (referred to as a representative pitch) that is most likely to be a pitch period in each subframe is obtained from the magnitude of the correlation value. To set the lag search range of multiple subframes. In setting the search range, an appropriate temporary pitch (referred to as a temporary pitch) having a small difference at the center of the search range is obtained using the representative pitches and correlation values of a plurality of subframes obtained by pitch analysis.

  Furthermore, since the lag search section is limited to a designated range before and after the provisional pitch obtained by setting the search range, an efficient search of the adaptive codebook is enabled. At that time, candidates for a short part of the lag are reduced, and a longer range of the lag is set wider, so that an appropriate search range in which good performance can be obtained can be set. Further, since the lag is searched in the range set by the setting of the search range during the adaptive codebook search, it is possible to perform encoding that can obtain a good decoded sound.

  Thus, according to the present embodiment, the temporary pitch of the second subframe is also present near the temporary pitch of the first subframe obtained by search range setting section 311. In the first subframe, Since the search range is narrowed down, the lag obtained as a result of the search does not go away from the temporary pitch. Therefore, when searching for the second subframe, the vicinity of the provisional pitch of the second subframe can be searched, and even in a non-stationary frame such as a case where speech starts from the second half of the frame, the first and second subframes are suitable. Lag search is possible, and speech coding can be performed without producing local noise.

  The speech encoding / decoding according to Embodiments 1 and 2 has been described as a speech encoding device / speech decoding device, but these speech encoding / decoding may be configured as software. For example, the speech encoding / decoding program may be stored in a ROM and operated according to instructions from the CPU according to the program. Further, the program, the adaptive codebook, and the stochastic codebook (pulse spreading codebook) are stored in a computer-readable storage medium, and the program, adaptive codebook, and stochastic codebook (pulse spreading code) of this storage medium are stored. (Book) may be recorded in the RAM of the computer and operated according to the program. Even in such a case, the same operations and effects as in the first and second embodiments are exhibited. Furthermore, the program in Embodiments 1 to 3 may be downloaded by a communication terminal and the program may be operated by the communication terminal.

  The first and second embodiments may be implemented individually or in combination.

The block diagram which shows the structure of the radio | wireless communication apparatus provided with the audio | voice coding apparatus of this invention. The block diagram which shows the structure of the audio | voice coding apparatus which concerns on Embodiment 1 of this invention. The block diagram which shows the structure of the gain calculating part in the audio | voice coding apparatus shown in FIG. The block diagram which shows the structure of the parameter encoding part in the audio | voice encoding apparatus shown in FIG. 1 is a block diagram showing a configuration of a speech decoding apparatus that decodes speech data encoded by a speech encoding apparatus according to Embodiment 1 of the present invention. Diagram for explaining adaptive codebook search FIG. 3 is a block diagram showing the configuration of a speech encoding apparatus according to Embodiment 2 of the present invention.

Explanation of symbols

102, 302 LPC analysis unit 103, 303 Adaptive codebook 104, 304 Probabilistic codebook 105, 305 Sound source creation unit 106, 306 Auditory weight LPC synthesis unit 107, 307 Audible weighting unit 108, 308 Gain calculation unit 109, 309 Parameter code Conversion unit 310 Pitch analysis unit 311 Search range setting unit 1091 Parameter calculation unit 1092 Encoding distortion calculation unit 1093 Comparison unit 1094 Vector codebook 1095 Prediction coefficient storage unit 1096 Decoded vector storage unit

Claims (4)

A CELP speech coding apparatus that performs coding by decomposing one frame into a plurality of subframes,
LPC synthesis means for obtaining a synthesized sound by filtering the adaptive sound source and the stochastic sound source stored in the adaptive code book and the stochastic code book using the LPC coefficient obtained from the input speech;
Gain calculation means for obtaining gains of the adaptive sound source and the stochastic sound source,
Adaptive encoding and stochastic excitation obtained using encoding distortion between the input speech and the synthesized speech, and parameter encoding means for performing vector quantization of the gain;
Before performing an adaptive codebook search of the first subframe of a frame, a correlation value is obtained by performing a pitch analysis of the plurality of subframes, and a value closest to the pitch period is calculated as a representative pitch using the correlation value Pitch analysis means to
Search range setting means for setting a search range of lags of the plurality of subframes based on the correlation value and the representative pitch period;
With
The adaptive codebook search is:
Performed on the search range of the lag,
A speech encoding apparatus characterized by that. The search range setting means includes
Using the correlation value and the representative pitch period, obtain a provisional pitch that is the center of the lag search range, and set a designated range before and after the provisional pitch as the lag search range.
The speech encoding apparatus according to claim 1. The search range setting means includes
Set a designated range before and after the provisional pitch so that candidates with a short lag with respect to the provisional pitch are fewer than candidates with a long lag.
The speech coding apparatus according to claim 2. CELP-type speech coding program that decomposes one frame into a plurality of subframes and performs coding; an adaptive codebook in which previously synthesized excitation signals are stored; a stochastic codebook in which multiple excitation vectors are stored; Is a computer-readable recording medium storing
The speech encoding program is
Before performing an adaptive codebook search of the first subframe of a frame, a correlation value is obtained by performing a pitch analysis of the plurality of subframes, and a value closest to the pitch period is calculated as a representative pitch using the correlation value And the steps to
A procedure for setting a search range of lags of the plurality of subframes based on the correlation value and the representative pitch period;
A procedure for performing the adaptive codebook search for the search range of the lag;
A procedure for obtaining a synthesized sound by filtering the adaptive sound source obtained by the adaptive code book search and the stochastic sound source stored in the stochastic code book using the LPC coefficient obtained from the input speech;
A procedure for obtaining gains of the adaptive sound source and the stochastic sound source;
Adaptive and stochastic sound sources determined using coding distortion between the input speech and the synthesized sound, and a procedure for performing vector quantization of the gain;
including,
A recording medium characterized by the above.

Images