Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L19/04—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using predictive techniques
  • G10L19/08—Determination or coding of the excitation function; Determination or coding of the long-term prediction parameters
  • G10L19/12—Determination or coding of the excitation function; Determination or coding of the long-term prediction parameters the excitation function being a code excitation, e.g. in code excited linear prediction [CELP] vocoders
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L2019/0001—Codebooks
  • G10L2019/0002—Codebook adaptations
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L2019/0001—Codebooks
  • G10L2019/0007—Codebook element generation
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L2019/0001—Codebooks
  • G10L2019/0013—Codebook search algorithms
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L2019/0001—Codebooks
  • G10L2019/0013—Codebook search algorithms
  • G10L2019/0014—Selection criteria for distances
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
  • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
  • G10L25/06—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being correlation coefficients

Definitions

• the present invention relates to a speech coding method for compressing information of a speech signal, and in particular, Ana 1 sis for coding at a transmission rate of 4 to 16 Kbps—by—Syntesis ⁇ A—b— S)
• the present invention relates to a speech coding method using vector quantization, that is, vector quantization that performs division by synthesis.
• A-b-S-type vector quantization such as Code — Excited Linear P rediction (CELP) encoders
• CELP Code — Excited Linear P rediction
• a prediction signal is applied to each code vector of a codebook to generate a reproduction signal, and the reproduction signal and the input sound signal are generated. It evaluates the error power between the two and determines the code vector number (index) with the least error and transmits it to the receiving side.
• the encoder using the A—b—S-type vector quantization method described above generates the ⁇ of the excitation signal of about 1000 patterns stored in the codebook.
• a linear prediction synthesis filter process is applied to each of the vectors, and one pattern that minimizes the error between each reproduced audio signal and the input audio signal to be encoded is obtained.
• a process of searching from a pattern of 0 is performed.
• the encoder since the encoder needs to be able to communicate immediately, it is necessary to perform the above search processing in real time. In this case, the search process must be performed continuously during a call at short time intervals, for example, 5 ms.
• this search process includes complicated operation operations such as filter operation and correlation operation, and the amount of operation required for these operation operations is, for example, 100 M multiplication 'addition / It is a huge thing of seconds.
• DSP Digital Signal Processor
• the present invention provides a speech code capable of dramatically reducing the amount of computation while maintaining the characteristics of an A-b-1 S-type vector quantization encoder of high quality and high efficiency.
• the purpose is to provide a chemical formula.
• a difference vector (hereinafter, referred to as a delta vector) between each code vector of the codebook and a previous code vector C, is provided.
• a delta vector the code vector group that generates the next code vector C n is stored in the codebook.
• n represents the order of the code vector group.
• Fig. 1 is a diagram for explaining the mechanism of speech generation
• Fig. 2 is a block diagram showing the general configuration of an A-b-S-type vector quantized speech encoder
• FIG. 3 is a block diagram specifically showing the codebook search processing portion in the configuration of FIG. 2,
• FIG. 4 is a diagram showing the basic concept of the present invention.
• FIG. 5 is a diagram simply showing the concept of the first embodiment based on the present invention
• FIG. 6 is a block diagram specifically showing a codebook search process part under the first embodiment
• FIG. 7 is a block diagram specifically showing another example of the codebook search process under the first embodiment
• FIG. 8 is a diagram showing another realization sequence of the autocorrelation calculation unit
• FIG. 9 is a block diagram specifically showing the codebook search process under the first embodiment with another example.
• FIG. 10 is a diagram showing another example of realizing the autocorrelation calculating unit
• FIG. 11 is a diagram showing a principle configuration of a second embodiment based on the present invention
• FIG. 12 is a diagram specifically showing the second embodiment of FIG. 11,
• FIG. 13 is a second embodiment.
• a diagram for explaining a delta vector of a tree-like array, which is a feature of Figures 14A, 14B and 14C show the distribution of code vectors (mode A, mode B and mode C) virtually formed in the codebook, respectively.
• FIGS. 15A, 15B and 15C are diagrams for explaining vector rearrangement based on the modified second embodiment.
• FIG. 16 is a diagram showing an example of a part of a codebook search process based on the modified second embodiment
• FIG. 17 is a diagram showing a successively optimized CELP type encoder
• FIG. 18 is a diagram showing a simultaneously optimized CELP type encoder
• FIG. 19 is a diagram showing an algorithm in FIG. 17,
• FIG. 20 is a diagram showing the algorithm in FIG. 18,
• FIG. 21A is a vector diagram schematically showing the gain optimization operation in the case of the sequential optimization CELP method,
• Figure 21B is a vector diagram that graphically illustrates the gain optimization operation in the case of the simultaneous optimization CELP method.
• Figure 21C is a vector diagram that graphically illustrates the gain optimization operation in the case of the Bitch orthogonal transformation optimization C ELP method.
• Fig. 22 is a diagram showing a coded orthogonal transform optimization CELP type encoder.
• FIG. 23 is a diagram specifically showing a part of the codebook search processing under the first embodiment with another example
• Figures 24A and 24B are vector diagrams to illustrate the Householder orthogonal transformation
• FIG. 25 is a diagram showing that the amount of calculation can be reduced by the first embodiment of the present invention
• FIG. 26 is a diagram showing that the amount of computation and the amount of memory can be reduced by the second embodiment according to the present invention.
• FIG. 1 is a diagram for explaining the mechanism of voice generation.
• Speech includes voiced and unvoiced sounds.
• Voiced sound is generated based on a pulsed sound source due to vibration of the vocal cords, and becomes a voice with the vocal tract characteristics of the mouth of the individual person added.
• An unvoiced sound is a sound that is emitted without shaking the vocal cords, and a simple Gaussian noise sequence becomes a sound source and becomes a voice through the vocal tract. Therefore, as shown in Fig. 1, the sound generation mechanism is based on the pulse source PSG, which is the source of voiced sound, the noise source NSG, which is the source of unvoiced sound, and the vocal tract of the signal output from each source (PSG, NSG). It can be modeled by a linear prediction synthesis filter LPCF that adds characteristics.
• PSG the pulse source
• NSG which is the source of unvoiced sound
• LPCF linear prediction synthesis filter
• the input speech is encoded using a code (data) that identifies the pulse period and the noise sequence of the noise source. can do.
• the pulse period of the pulse sound source is identified based on the periodicity of the input speech signal using the adaptive codebook, and a pulse train having the period is input to the linear predictive synthesis filter to perform a filter operation process.
• the obtained filter operation result is subtracted from the input audio signal to remove the periodic component.
• a plurality of noise sequences each noise The sequence is represented by an N-dimensional predetermined code vector
• the reproduced signal vector obtained by applying a synthesis filter to each code vector and the input signal from which the periodic component has been removed are prepared. If one code vector that minimizes the error from the vector (N-dimensional vector) is found, the speech is encoded using the code (data) that specifies the period and the code vector. Then, this data is transmitted to the receiving side, and the original audio (input audio signal) is reproduced. This data is highly compressed information.
• Figure 2 is a block diagram showing the general configuration of an A-b-S-type vector quantized speech coder.
• reference numeral 1 denotes a plurality of randomly generated noise sequences C, for example, 10 2 4 types.
• Each noise sequence is represented by an N-dimensional code vector
• 2 is a gain unit for gain g
• 3 is a synthetic filter that simulates vocal tract characteristics in the output of the amplification unit.
• 4 is a linear prediction synthesis filter that performs arithmetic processing on the filter
• 4 is an error generator that outputs the error between the reproduced signal vector and the input signal vector output from the linear prediction synthesis filter 3
• 5 is the error evaluation.
• a noise sequence that minimizes the error
• A—b—S vector quantization differs from normal vector quantization in that each code vector (C) in the noise codebook 1 is multiplied by the optimal gain g, and then linear predictive synthesis is performed.
• Filter 3 performs filter processing, and the reproduced signal vector obtained by the filter processing
• (E) is calculated by the error generator 4 and the error signal is calculated by the error power evaluator 5.
• the noise codebook 1 is searched using the power of the noise as an evaluation function (distance scale) to find a noise sequence (code vector) that minimizes the error power, and the noise sequence (code vector) is obtained.
• the input audio signal is encoded using a code that specifies A is the auditory weighting matrix ⁇ The above error power is
• Equation (1) The optimal code vector C and gain g are determined so as to minimize the error power shown in equation (1). Since the power varies depending on the volume of the voice, the gain g is optimized to match the power of the reproduced signal g AC with the power of the input audio signal X. The optimal gain can be obtained by partially differentiating equation (1) with g and setting it to 0. That is,
• R xc and the auto-correlation of the combined filter output AC are R cc, and the cross-correlation and auto-correlation are
• R CC (AC) T (AC) (5)
• T represents a transpose (transp 0 se) matrix. Since the code vector C that minimizes the error power E in Eq. (3) maximizes the second term on the right side of the equation, the code vector C is
• a r g max is the maximum of a r g u m e n t
• the optimal gain can be obtained from Eq. (2) using the cross-correlation and autocorrelation that satisfy Eq. (6).
• FIG. 3 is a block diagram specifically showing a codebook search processing portion in the configuration of FIG.
• it is a configuration diagram of a noise codebook search processing portion that calculates a noise sequence (code vector) that minimizes error power and encodes an input signal.
• Reference number 1 is a noise codebook that records M types (size M) of noise sequences C (each noise sequence is represented by an N-dimensional code vector), and 3 is a vocal tract characteristic.
• simulated the full I filter processing performing a linear predictive synthesis off I filter minute folding order N P (LPC off I filter).
• the illustration of the G amplifier 2 in FIG. 2 is omitted.
• 7 is a square calculator that calculates the square of the cross-correlation R xc
• 9 is a divider that computes R xc 2 ZR CC
• 10 is a noise sequence that maximizes R XC 2 ZR cc , in other words, a noise sequence that minimizes error power
• It is an error power evaluation determination unit that determines a code vector and outputs a code specifying the code vector.
• the main ones are (1) filter processing on code vector C, (2) calculation processing of cross-correlation R xc , and (3) calculation processing of auto-correlation R cc
• the three assuming that the order of the LPC filter 3 is N P and the dimension of the vector quantization (code vector) is N, for one code vector, the above (1) to (1) 3)
• the amount of computation required for each term is NP ⁇ N, N and N. Therefore, one code base-vector per calculation amount required for the codebook search (N P + 2) - a N.
• FIG. 4 is a diagram showing the basic concept of the present invention.
• M noise sequences of N dimensions each are the code vector C.
• C, C 2 C 3 , C,-C n the code vectors.
• FIG. 5 is a diagram simply showing the concept of the first embodiment based on the present invention.
• the next code vector for example, the i-th code vector C i is the same as the previous code vector. It is calculated by adding the delta vector ⁇ C i to the torque, that is, the code vector C i—. At this time, the delta vector ⁇ C i needs to be as simple as possible as described above.
• the columns of black circles drawn along the horizontal axis in the columns of C i—,, AC i and C i in FIG. 5 are N (N samples) in the case of an N-dimensional code vector. This corresponds to each sample point in the noise train waveform.
• Fig. 5 shows an example consisting of only four significant sample data ⁇ 1, ⁇ 2, ⁇ 3 and ⁇ 4 as the delta vector ⁇ C i, which is extremely simple c
• the code vectors of the random codebook 1 may be rearranged so that they are simple. In other words, taking any two adjacent code vectors (C i-!, C i), the delta vector between them Each code vector is arranged into an original table so that (m Ci) becomes a simple vector of several sample data as shown in Fig. 5.
• code vectors, C i in Fig. 5 are examples using code vectors that have already been sparsed, that is, pre-processed to provide many codes with zero sample values. I have. However, the code vector sparsification technology is well known.
• a delta vector group in which the difference between any two adjacent code vectors C i-, and C i becomes a simple delta vector AC i is sequentially determined. It is stored in the delta vector codebook i 1 (described later).
• FIG. 6 is a block diagram specifically showing a part of the codebook search processing under the first embodiment. Basically, it corresponds to the configuration of FIG. 3 described above, but FIG. 6 shows an example in which the present invention is applied to a well-known successively optimized CELP type speech encoder. Therefore, an auditory weighted pitch prediction error signal vector AY is shown in place of the input speech signal X (FIG. 3), but has no effect on the description of the present invention.
• the calculation means 19 is shown, This is a preprocessing means associated with the transfer of the linear prediction synthesis filter 3 from the position shown in FIG. 3 to the position shown in FIG. 6-an important element for understanding the present invention. Absent.
• the part that generates the cross-correlation R xc in FIG. 3 corresponds to the cross-correlation calculator 12 in FIG. 6, and the part that generates the auto-correlation R cc in FIG. This is the autocorrelation calculation unit 13 in FIG.
• the cross-correlation calculator 12 includes a cyclic addition unit 20 for realizing the present invention, a power addition unit 14 and a delay unit 15.
• the above cyclic addition means 20 for realizing the present invention in the autocorrelation function part 13 is shown as an addition part 16 and a delay part 1.
• each code vector is similar to the noise codebook 1 in Fig. 3.
• the perceptually weighted pinch prediction error signal vector AY is converted into AT AY by the arithmetic means 21 and
• the delta vector ⁇ C of the delta vector codebook 11 is directly supplied to the cross-correlation calculator 12 for multiplication and the correlation value of the previous image.
• the autocorrelation calculation unit 13 generates the code vector C by cyclically adding the delta vector AC to the front code vector Ci—,
• the autocorrelation value (AC) T AC of the code vector AC after auditory weighting reproduction is obtained and given to the evaluation unit 5.
• the cross-correlation calculation unit 12 and the auto-correlation calculation unit 13 need only be multiplied by the sparsed delta vector, the calculation amount can be reduced.
• FIG. 7 is a block diagram specifically showing another example of the codebook search process under the first embodiment, which is applied to a well-known simultaneous optimization CELP type speech encoder. Show the case. In this case, the present invention can be applied without any trouble. Also in this figure, the first and second arithmetic means 19-1 and 191-2 are not directly related to the present invention. Since the cross-correlation calculation unit (12) is processed in parallel while being divided into an input speech system and a pitch P (period described above) system, the first and second cross-correlation calculation units 1 2 — 1 and 1 2 — Two.
• FIG. 8 is a diagram showing another example of realizing the autocorrelation unit, and the autocorrelation calculation unit 13 shown in FIGS. 6 and 7 can be realized with other configurations.
• the arithmetic unit 21 shown here is such that in FIG. 6 and FIG. 7, the multiplications required in the combining filter 3 and the auto-correlation calculation unit 8 can be completed by one multiplication.
• the previous code vector C i—! In addition to storing the min value A ⁇ A of the auditory weighting matrix A and calculating the delta vector ⁇ C i, the addition unit 16 and the delay unit 17 (cyclic addition means)
• FIG. 9 is a block diagram specifically showing the codebook search processing part under the first embodiment with another example.
• FIG. Corresponding to the configuration in the figure, Fig. 9 shows an example of application to a pitch orthogonal transform optimization CEL ⁇ -type speech coder.
• a block placed after the arithmetic means 19 ' is shown.
• Reference numeral 22 denotes a time-reversal orthogonal transform unit, which calculates a time-reversal perceptually weighted input voice signal vector A T AX from the perceptually weighted input voice signal vector AX by an arithmetic means 19 ′.
• the time-reversed orthogonal transform unit 22 calculates the time-reversed perceptually weighted orthogonally transformed input speech signal vector (AH) T AX for the pitch prediction residual vector AP weighted at the optimal time. Is what you do. However, these arithmetic means 19 ′ and the time-reversal orthogonal transform section 22 have no direct relation to the essence of the present invention.
• the cross-correlation calculation unit 12 performs multiplication with the delta vector C and cyclic addition as in the case of FIGS. 6 and 7, and evaluates the correlation value (AHC) T AX. This is a matrix that represents the e H m.
• the autocorrelation calculation unit 13 cyclically adds the delta vectors ⁇ C i of the delta vector codebook 11 with the addition unit 16 and the delay unit 17 to obtain a code vector.
• Generate C i and calculate the code vector AHC AC ', which is the auditory weighted orthogonal transform of the auditory weighted (A) pitch prediction residual vector AP at the optimal time.
• the autocorrelation value (AHC) T AHC (AC ') T AC' of the code vector AHC subjected to the auditory weighted orthogonal transformation.
• the amount of calculation by the delta vector can be reduced in the same manner when performing the orthogonal orthogonal transformation optimization.
• FIG. 10 is a diagram showing another example of realizing the autocorrelation operation unit.
• the autocorrelation operation unit 13 shown in FIG. 9 can also be realized by another configuration. This corresponds to the configuration shown in FIG.
• the arithmetic unit 23 shown here is such that in FIG. 9, the multiplications required in the synthesis filter (A H) 3 ′ and the auto-correlation operation unit 8 can be completed by one multiplication.
• the arithmetic unit 23 stores the previous code vector! And the correlation value (AH) T AH of the perceptual weighting matrix AH subjected to the orthogonal transformation, and stores the correlation value with the delta vector Ci.
• cyclic addition is performed by the adder 16 and the delay 17 so that (AHC i) T AHC i
• the delta base-vector A d, delta C 2 by linearly accumulate ⁇ C 3 ⁇ ⁇ ⁇ , virtually, code Dobeku stored in conventional noise codebook 1 Tor C i, C 2 , C
• a delta vector group is generated under a special rule, thereby significantly reducing the amount of calculation required for the codebook search process.
• the second embodiment also has an advantage that the memory capacity in the delta vector codebook 11 can be significantly reduced.
• the second embodiment will be described.
• FIG. 11 is a diagram showing a principle configuration of a second embodiment based on the present invention.
• the concept of the second embodiment is schematically illustrated in the upper half of FIG. 11, and a delta for generating a virtually formed code vector of, for example, 124 patterns is provided.
• the vectors are arranged in a tree structure with a certain rule with a certain polarity. As a result, only (L_1) filter operations and correlation operations
• the above-mentioned autocorrelation calculation unit. 10 is the already-determined noise sequence (code vector) that maximizes R xc 2 ZR cc , in other words, minimizes the error power.
• the error power evaluation decision unit 30 described above is a speech encoding unit that encodes the input speech signal with data (code) that specifies a noise sequence (code vector) that minimizes the error power. The operation of this encoder is as follows.
• the initial vector (:. And the (L1-1.)) kinds of delta noise sequences ie, the delta vector C, ⁇ ⁇ C! may be stored in-vector codebook 1 1, delta base click preparative Le ⁇ C ⁇ ⁇ C L -! , the initial base click preparative Le respectively
• the code vector C of (2 '° -1) types of noise sequences in a tree structure.
• Add the vector and 2 1 This represents the code vector (:. ⁇ C10Z3 ) of the pattern in this way.
• the initial vector (:. And (L_1) kinds of delta vectors are stored in the delta vector codebook 11 1.
• the error power evaluation determination unit 10 determines a noise sequence (code vector) that minimizes the error power
• the speech coding unit 30 outputs a code specifying the code vector.
• Rxc ( 2K + Z ) is
• Rxc (2k + 1) Rxc (k) + ⁇ ⁇ (A ⁇ C i)
• Rxc ⁇ zk + 2) Rxc (k) -X T (AA Ci) (13)
• the cross-correlation calculation unit 12 generates the cross-correlation R xc ( ⁇ > To calculate the current cross-correlation R xc (2k + n , R xc (2k + z> ).
• the self-correlation calculation unit 13 calculates the current cross-correlation R cc (2k + 1 ) and R cc (Zk + 2> using the auto-correlation R cc (k > one layer before.
• the autocorrelation R cc is calculated as the sum of L of the filter output AC of the initial vector and the filter output AAC of (L-1) kinds of delta vectors, ⁇ k. autocorrelation ⁇ ⁇ , ( ⁇ ⁇ C!) 2 ⁇ (AC.) (AAC L - 1) and 2, each full I filter output AC, AAC ⁇ AAC L -.! ! between (L 2 - 1) / Can be calculated using two types of cross-correlation.
• FIG. 12 is a diagram specifically showing the second embodiment of FIG. 11, and as described above, 11 is an initial vector (:. And (L) representing one reference noise sequence.
• An N-dimensional vector obtained by encoding the amplitude of the sampled noise is shown in Fig. 3.
• the above-mentioned linear prediction synthesis filter (LPC filter) that performs filter operation processing that simulates vocal tract characteristics a filter), is composed of IIR (I nfinite I mulse R esponse ) type full I filter of order N P, square Conclusions Li click scan a and U-NXN
• the matrix operation of the vector C is performed, and the synthesis process is performed on the code vector C.
• the N P coefficients of the IIR filter are based on the input audio signal X.
• the correlation coefficient exists between adjacent samples of the input audio signal, so that the correlation coefficient between the samples is determined, and the correlation coefficient is referred to as the Percoll coefficient.
• Calculated partial autocorrelation coefficient and determines Alpha coefficient of IIR filter from the Percoll coefficient
• a square matrix A of NXN is created by using the impulse response train of the filter, and a synthesis filter process is performed on the code vector C.
• 31 1 is a filter operation processing for the initial vector (:. And the delta vector C i to A C L —, which represent the delta noise sequence representing the reference noise sequence.
• Filter output AC obtained by applying> AA C, ⁇ A A C L —
• the operation unit 38 includes a square operation unit 7 and a division unit 9, and
• 10 is an error power evaluation determination unit that determines a noise sequence (code vector) that maximizes Rxc 2 / R cc , in other words, that minimizes error power.
• This is a speech coding unit that codes the input speech signal with a code that specifies the noise sequence (code vector) that minimizes the error power.
• FIG. 13 is a diagram for explaining a delta vector of a tree-like array characteristic of the second embodiment.
• Synthesis full I filter operation output for the code base-vector C 2k + C zk + 2 AC 2k + 1, AC Zk + 2 includes a synthesis filter operation output A c k for a hierarchy previous code base click Honoré c k Using the filter output AA Ci of this delta vector,
• the conventional codec of all noise sequences is C.
• the cross-correlation calculation unit 12 can calculate the current cross-correlation R xc ( : 2 z k + I ') R xc (Zk + 2> using the cross-correlation R xc (k) one layer before. In this way, the filter output and input voice for the code vector of the entire noise sequence can be obtained simply by performing the cross-correlation calculation of the second term on the right side of the above equations (20) and (21). It is possible to calculate the cross-correlation with the signal X. That is, the conventional cross-correlation
• multiply-accumulate is required twice.
• reference numeral 6 denotes a multiplier for calculating the second term ⁇ ⁇ ( ⁇ C i) on the right side of the above equations (20) and (21).
• 3 5 is a polarity imparting part that generates +1, — 1,
• 3 6 is a multiplication part that multiplies the polarity ⁇ 1 to give the polarity to the second term on the right side
• 15 is the previous cross-correlation R xc (k> (the first term on the right side) is a delay unit that stores and delays for a predetermined time
• 14 is the sum of the first and second terms on the right side of the above equations (20) and (21).
• the combined filter operation output AC zk + 1 and AC 2k + 2 are calculated using the combined filter operation output AC k of the previous layer and the filter output AAC i of the current delta vector (18).
• the self-phase ⁇ R cc for the code vector of each noise sequence is expressed by the following equation. That is,
• a C z A C o _ A A C i
• R cc (1) (AC o) T (AC.)
• Rcc (2) (AC.) T (AC. :) ten
• R (6) (A C.) T (A C o) +
• the auto-correlation ( ⁇ Ci) ⁇ ( ⁇ ⁇ C i) of the current A ⁇ C is added to the self-correlation ⁇ R cc ⁇ k) one level before, and the A
• the cross-correlation R cc ⁇ 21 ⁇ + ", R cc ( 2k + 2) can be calculated by adding the cross-correlation of Aum Ci to ⁇ C i-, while switching the polarity (10, 1). Then, the autocorrelation R cc is set to the filter output AC of the initial vector.
• the correlation is given polarity (+, —)
• the cross-correlation synthesizing unit 16 adds the auto-correlation R cc ⁇ k> one level before and the auto-correlation (AAC i) ⁇ (AAC i) and the cross-correlation (2 3) , (24) is the above-mentioned addition unit, and 17 is the above-mentioned delay unit that stores and delays the autocorrelation R cc (k) of the immediately preceding layer for a predetermined time.
• a C. , AA d A A C L- ⁇ L-I O) are obtained and stored in the storage unit 31.
• i is set to 0, and the cross-correlation operation unit 12 performs cross-correlation.
• the error power evaluation determining unit 10 compares the calculated value F (X, C) with the maximum value F max (initial value is 0) of F (X, C) so far, and calculates F (X, C )> if F max, F (X, C ) ⁇ as F max updates the F max, 1 single code giving F max The code up to that point is updated with the code (index) that specifies the vector.
• the error power evaluation determining unit 10 compares the calculated value F (X, C) with the maximum value F max (initial value is 0) of F (X, C) so far, and calculates F (X, C )> F max , F max is updated as F (X, C) ⁇ F max , and a code (index) specifying one vector that gives the F max is used until that time. Update the sign of.
• the calculation unit 38 calculates the above (14!
• the error power evaluation deciding unit 10 compares the calculated value F (X, C) calculated in the same manner with the maximum value F max (initial value is 0) of the previous F (X, C), and calculates F (X, C)> If Fmax , F (X, C) ⁇ Fmax is updated, and a code (index) for specifying one code vector that gives the Fmax is provided. And update the previous sign.
• the voice coding unit 30 uses the latest code (index) stored in the error power evaluation determining unit 10 as voice coding information for the input voice signal. Output as
• C is the initial vector.
• the ninth delta vector which is the component of the lowest hierarchy, the component of ⁇ C 9, is contained in half of the vector, while the component of C 9 is included in all the vectors. It is only possible (see Fig. 13). That is, the contribution of each delta vector to the structure of the codebook 11 is not uniform, and the delta vectors that constitute the upper layer of the tree structure array, for example, the initial vector C o, the first It can be said that the delta vector ⁇ C, etc., is included as a component in many code vectors and determines the codebook distribution mode.
• FIG. 14A, 14B and 14C show the distribution of code vectors (mode A, mode B and mode!) Virtually formed in the codebook, respectively.
• FIG. For example 3 Tsunobeku preparative Le, i.e. (:., AC, and when considered ⁇ C 2, this There are six distributions of these vectors (modes A to F).
• Modes A to F are shown in Figures 14A to 14C, respectively.
• e x, e y and e z represent the x-axis constituting the three-dimensional, y-axis and z-axis in each direction of the unit base-vector of each.
• the remaining modes, D> E and F are equivalent when unit vectors are assigned to each vector as follows:
• the order of the total L vectors given as the initial vector C Q and the delta vector AC is rearranged, so that the code book 11 Adjust the mode of distribution of the formed code vector. That is, the properties of the codebook are changed.
• the above code vector distribution mode is encoded. Adjust according to the characteristics of the input audio signal to be applied. Thereby, the quality of the reproduced sound is further improved.
• the vectors are rearranged for each frame according to the characteristics of the linear predictive synthesis (LPC) filter 3.
• LPC linear predictive synthesis
• the power of the filter output of the vector obtained by performing the linear prediction synthesis filter processing on the initial vector and each delta vector The power is evaluated, and this power is the vector that has the largest increase compared to the power before the filter processing.
• the initial vector, the first delta vector, the second delta vector, and so on Rearranging is a specific example.
• these vectors are transformed in advance so that the initial vector and each delta vector are orthogonal to each other after the linear prediction synthesis filter processing.
• the vectors virtually formed in the codebook 11 can be uniformly distributed on the hyperplane (HyperP1ane).
• it is preferable that the power of the initial vector and the power of each delta vector is normalized (norma1ize). As a result, the power of the filter output of each vector can be rearranged by simply comparing the powers.
• the inter-code distance (vector Euclidean distance) between vectors belonging to the upper layer of the pattern vector array is the inter-code distance between vectors belonging to the lower layer.
• the code of the audio coded information is allocated so that it becomes larger than the distance. This focuses on the fact that the vectors belonging to higher layers (such as the initial vector and the first delta vector) have a greater effect on the quality of the reproduced sound obtained by decoding on the receiving side. It was done. As a result, even if a transmission error occurs on the transmission path to the receiving side, it is possible to minimize the deterioration of the quality of the reproduced sound.
• FIGS. 15A, 15B and 15C are diagrams for explaining the rearrangement of vectors based on the modified second embodiment.
• a sphere centered at the origin of coordinates is the space of all base-vector defined by the unit base-vector e x, e y and e z, if the initial base-vector C .
• the allocated a unit base click preparative Le e, also when the first delta base-vector ⁇ C, and the second base click preparative Le ⁇ C 2 base units-vector e y and e z is that vignetting with split, these
• the characteristic A of the linear prediction synthesis filter 3 shows a different amplitude amplification characteristic for each vector which is a constituent element of the delta vector codebook 11. It is better to make all the virtually formed vectors non-uniform than to distribute them evenly over the entire vector space. Therefore, it is examined which vector component is amplified most in any direction, and if the distribution of the vector component in that direction is increased, the vector can be efficiently added to the codebook 11. Can be stored, resulting in improved quantization characteristics of the audio signal.
• the distribution of tree-structured delta vectors is biased, but by changing the order of the delta vectors, the properties of the codebook 11 can be changed.
• FIG. 16 is a diagram showing an example of a codebook search process based on the modified second embodiment, which is shown in FIGS. 15A, 15B and 15C.
• An example of performing the expressed sorting will be described. This is equivalent to a modification of the configuration of FIG. 12 (second modified embodiment) described above, and is different from the configuration of FIG. 12 in FIG.
• the power estimator 41 adds the initial vector and each delta after the filter processing by the linear prediction synthesis filter 3. Evaluate the vector power.
• the sorting unit 42 the order of the vectors is rearranged based on the magnitude of the amplitude amplification rate of each vector obtained as a result of this evaluation.
• the power evaluation unit 41 and the sorting unit 42 will be described below with reference to FIGS. 14A to 14C and FIGS. 15A to 15C described above. .
• AC. A Calculate the power (power) of ⁇ C !, AAC 2 ).
• the power of each vector is normalized in advance (see (1) below)
• the power after the filter processing can be directly compared, and the amplitude for each vector can be directly used as it is.
• the amplification rates will be compared (see (2) below).
• the amplitude amplification factor of each vector by the synthesis filter (A) is received from the power evaluation section 41, and the vectors are sorted (sorted) in descending order of the amplification factor.
• Subsequent encoding processing uses the tree-structured delta codebook 11 composed of the obtained delta vectors, and the tree-structured delta code in FIG.
• the sorting process for the case shown in Fig. 15A to 15C is shown below.
• the second embodiment and the modified second embodiment described above are different from the first embodiment described above.
• the present invention can be applied to any of a sequential optimization CELP-type speech coder and a simultaneous optimization CELP-type speech coder or a pitch orthogonal transform optimization CELP-type speech coder.
• the method of application is the cyclic addition means 20 (14, 15; 16, 17; 1-1, 1 5-1; 14-2, 15-2) described in detail in the first embodiment. ).
• FIG. 17 is a diagram showing a successively optimized CELP type encoder
• FIG. 18 is a diagram showing a simultaneously optimized CELP type encoder. The same components as those already described are indicated by the same reference numerals or symbols.
• adaptive codebook 101 stores an N-dimensional pitch prediction residual vector corresponding to N samples whose Bitch period is delayed by one sample
• codebook 1 also stores
• the code vector generated using the N-dimensional noise sequence corresponding to the same N samples is set in advance for 2 m patterns.
• the sample data whose amplitude is equal to or smaller than a certain value for example, N / 4 samples of N samples
• Such a codebook is called a sparse codebook.
• A 1 / A '( Z ) (where A' (z) indicates an auditory weighted linear prediction analysis filter) for each pitch prediction residual vector P of the adaptive codebook 101.
• Indicated auditory weighted linear prediction sum The pitch prediction vector AP generated by the auditory weighting by the synthesis filter 103 is multiplied by the gain b by the amplifier 105 to generate a pitch prediction reproduction signal vector bAP.
• an auditory weighting filter represented by the pitch prediction reproduction signal vector bAP and A (z) / A '(z) (where A (z) indicates a linear prediction analysis filter).
• the pitch prediction error signal obtained by the subtraction unit 108 is obtained by calculating the perceptually weighted bitch prediction error signal vector AY between the input audio signal vector AX and the input speech signal vector AX weighted in 107.
• the evaluation unit 110 sets the following equation for each frame so that the power of the vector AY becomes the minimum value.
• each code vector C of codebook 1 is similarly amplified by the linear predictive synthesis filter 3 to the code vector AC after auditory weighting generated and reproduced by auditory weighting.
• the gain g is multiplied by 2 to generate a linear predictive reproduction signal vector g AC.
• the position of the amplifier 2 may be located before the filter 3.
• Evaluator 5 selects the optimal code vector C from codebook 1 for each frame and minimizes it so that it becomes the minimum value by 2 (2 6). What Select gain g.
• suitable, codebook 1 0 1 adaptation b the AP + g AC determined by the summing unit 1 1 2, which perceptual weighting linear prediction minute ⁇ full Note1 ( ⁇ '( ⁇ )) 1 1
• This is performed by dividing into bP + gC in 3 and delaying it by one frame by the delay unit 114 and storing it as an adaptive codebook (pitch prediction codebook) for the next frame.
• the error signal vector E from the signal vector is obtained as described above, and the evaluation unit 5 selects the code vector C that minimizes the power of this vector E from codebook 1.
• the optimal gains b and g are simultaneously selected and controlled.
• the adaptation of adaptive codebook 101 in this case is performed in the same manner for AX ′ corresponding to the output of adder 112 in FIG.
• the gain bg conceptually shown in FIGS. 17 and 18 is actually different from each other as shown in FIGS. 19 and 20.
• code vector C of codebook 1 is optimized.
• the pitch prediction error signal vector AY is multiplied by the code vector AC obtained through the auditory weighting linear prediction synthesis filter 3 by the code vector C of the codebook 1 Multiplied by 6 to generate a correlation value (AC) ⁇ AY between them, and the code vector after auditory weighting reproduction
• V (AP) T AP (AC) T AC
• the input speech signal vector AX weighted by the auditory sense and the code vector AC of the codebook 1 by the code vector AC obtained through the auditory weighted linear predictive synthesis filter 3 are used.
• the evaluation unit 5 calculates the auditory weighting input audio signal vector AX by the above equation (29).
• the optimal code vector C and the gains b and g that minimize the power of the error signal vector E are selected at the same time. Again, minimizing the power of vector E is dependent on the relative value ratio,
• the total amount of computation is smaller than in the case of the simultaneous optimized CELP system, but the quality of the encoded speech is degraded.
• Figure 21A is a vector diagram that schematically illustrates the gain optimization operation in the case of the sequential optimization CELP method
• Figure 21B is a graph that illustrates the gain optimization operation in the case of the simultaneous optimization CELP method
• FIG. 21C is a vector diagram schematically illustrating the gain optimization operation in the case of the pitch orthogonal transform optimization CELP method.
• each code vector C of Codebook 1 is reproduced as an auditory weighting orthogonal to the pitch prediction residual vector AP at the optimal time among the pitch estimation residual vectors weighted in the auditory weight. It does this by converting it to the code vector AC '.
• the auditory weighted pitch prediction playback vector bAP is extracted from Codebook 1 and subjected to the auditory weighting matrix A.
• the non-orthogonal code vectors AC are not orthogonal, which causes the quantization error ⁇ to increase in the successive optimization method as in the 2 1 AH.
• the code vector AC is orthogonally transformed to the code vector AC ′ orthogonal to the auditory weighted Bitch prediction residual vector AP by a known method, the sequential optimization of FIG. In the CELP method, the quantization error can be reduced to the same level as in the simultaneous optimization method.
• the code vector AC ′ thus obtained is multiplied by a gain g to generate a linear predictive reproduction signal g AC, and the linear predictive reproduction signal g AC ′ and the auditory weighted input speech Traffic light
• the evaluation unit 5 selects a code vector from the codebook 1 that minimizes the linear prediction error signal vector E from the vector AX, and selects the gain g.
• FIG. 23 is a diagram specifically showing the codebook search processing part under the first embodiment with another example, and the above-described pitch orthogonal transform optimization CELP type speech encoder Shows the case when applied to. In this case, the present invention can be applied without any trouble.
• FIG. 23 shows an example in which the configuration shown in FIG. 9 is combined with the self-correlation operation unit 13 in FIG.
• the arithmetic means 19 ′ shown in FIG. 9 can be constituted by a transposition matrix A ⁇ similarly to the arithmetic means 19 in FIG. 6, but in this example, a time inversion type filter is used. It is composed of Ruta.
• the autocorrelation calculation section 60 in this figure is composed of calculation sections 60a to 60e.
• This vector V is input with a vector D orthogonal to all delta vectors ⁇ C of the delta vector codebook 11 1, and is subjected to an auditory weighting filter (A) process.
• A auditory weighting filter
• H T A performs Of this Chinobeku Torr B and u B is house holder orthogonal transform and time-reversed with respect to A T AX output from the arithmetic unit 7 0 is sent to the time-reversed orthogonal transformation section 7 1 T AX-(AH) Generates T AX.
• the D-direction vector obtained in this way is taken in the D-direction, that is, in the opposite direction— (iV! / IDI) D as shown in the figure.
• the vector B V—
• vector C 'orthogonal to V is obtained. That is,
• the operation configuration is as shown in the figure.
• the portion indicated by O represents the vector operation
• the portion indicated by ⁇ represents the scalar operation.
• Gram-Schmidt method is also known as an orthogonal transformation method.
• R XC (AC) T (AH) T A
• AHC T AX is obtained by cyclically adding these values to the cyclic addition unit 67 (cyclic addition means 20). Will be sent to the price department 5.
• This self-correlation value can be obtained and sent to the evaluation unit 5.
• FIG. 25 is a diagram showing that the amount of calculation can be reduced by the first embodiment according to the present invention.
• (A) ⁇ in the figure is the case of the successive optimization CELP type encoder, and (1) Conventional 4-5 sparse codebook,
• N is the number of samples
• N p is the order of the filter 3.
• the total computational complexity is 4 32 K in the conventional example (1) and 8 4 K in the conventional example (2), as shown at the right end of the figure. It is multiply-accumulate.
• the product-sum operation is significantly reduced to 28 K.
• FIG. 26 is a diagram showing that the amount of calculation can be reduced and the amount of memory can be eliminated according to the second embodiment of the present invention.
• Column (a) of this figure shows the amount of computation, and column (b) shows the amount of noise codebook G) memory.
• L L
• This L is the same as the number of layers 1, 2> 3 ... shown in the upper part of FIG.
• the total amount of calculation shows that the conventional method is 480
• the total amount of filter processing, autocorrelation, and cross-correlation, which occupy the majority of the amount of operation, is drastically reduced, similarly to the values shown in FIG. .
• a difference vector between adjacent code vectors (delta vector) is adopted as a code vector to be stored in the random codebook. I do. As a result, the amount of calculation is further reduced than before.
• the first embodiment is further improved
• the quality of the reproduced sound can be further improved.
• the present invention can be applied to a transmission system in a cellular phone or a car phone, particularly to a speech encoder for transmitting input voice as digital data to a reception system.

Landscapes

• Engineering & Computer Science (AREA)
• Computational Linguistics (AREA)
• Signal Processing (AREA)
• Health & Medical Sciences (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Human Computer Interaction (AREA)
• Physics & Mathematics (AREA)
• Acoustics & Sound (AREA)
• Multimedia (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)
• Transmission Systems Not Characterized By The Medium Used For Transmission (AREA)

Priority Applications (4)

Applications Claiming Priority (4)

Publications (1)

Family

ID=26463564

Family Applications (1)

Country Status (6)

Cited By (1)

Families Citing this family (36)

Citations (3)

Family Cites Families (5)

• 1991
  • 1991-09-17 WO PCT/JP1991/001235 patent/WO1992005541A1/ja not_active Ceased
  • 1991-09-17 EP EP91915981A patent/EP0500961B1/en not_active Expired - Lifetime
  • 1991-09-17 JP JP03515016A patent/JP3112681B2/ja not_active Expired - Fee Related
  • 1991-09-17 DE DE69129329T patent/DE69129329T2/de not_active Expired - Fee Related
  • 1991-09-17 US US07/856,221 patent/US5323486A/en not_active Expired - Lifetime
  • 1991-09-17 CA CA002068526A patent/CA2068526C/en not_active Expired - Fee Related

Patent Citations (3)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events