JP3579276B2 - Audio encoding / decoding method - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a low coding rate speech encoding / decoding method used for digital telephones, voice memos, and the like.
[0002]
[Prior art]
2. Description of the Related Art In recent years, as a coding technique for compressing voice and musical sound into a small amount of information for transmission and storage on a mobile phone or the Internet, a CELP method (Code Excited Linear Prediction (MR Schroeder and BS Atal, "Code Excited Linear Prediction (CELP): High Quality Speech at Very Low Bit Rates," Proc. ICASSP, pp. 937-940, 1985 (Literature 1) and W. S. Kj. Improved Speech Quality and Efficient Vector Quantification in SELP, "Proc. ICASS P, pp. 155-158, 1988 (Reference 2)).
[0003]
CELP is a coding method based on linear prediction analysis, and an input speech signal is divided into a linear prediction coefficient representing phoneme information and a prediction residual signal representing a pitch or the like by linear prediction analysis. A recursive digital filter called a synthesis filter is configured based on the linear prediction coefficients, and the original input audio signal can be restored by inputting a prediction residual signal as a drive signal to the synthesis filter.
[0004]
In order to perform encoding at a low rate, it is necessary to encode a linear prediction coefficient, which is synthesis filter information representing characteristics of a synthesis filter, and a prediction residual signal, which is a drive signal for driving the synthesis filter, with a smaller amount of information. is there. In the CELP method, a signal obtained by encoding a prediction residual signal is generated as a drive signal by multiplying two types of signals, that is, a pitch vector and a noise vector, by an appropriate gain and then adding the signals. A method of generating a pitch vector is described in, for example, Reference 2.
[0005]
In addition to the method of Reference 2, a method of using a fixed code vector in a voice rising part (onset) has been proposed, but in the present invention, these are collectively referred to as a pitch vector.
The noise vector is usually generated by storing a large number of candidates in a random codebook and selecting an optimum one from the stored candidates. As a noise vector search method, all the noise vectors are added to the pitch vector, then a synthetic speech signal is generated through a synthesis filter, and the distortion of the synthesized speech signal with respect to the input speech signal is evaluated. A method of selecting a noise vector for generating an audio signal is used. Therefore, how to efficiently store the noise vector in the noise codebook is an important point of the CELP method.
[0006]
The Algebraic Codebook (JP P. Adoul et al, "Fast CELP Coding based on algebraic codes", Proc. ICASPSP '87, pp. 1957-1960 (pp. 1957-1960) describes the noise vector as a pulse. It is a simple structure expressed only by presence / absence and polarity (+,-). The algebraic structure codebook has features that it is not necessary to store a code vector and the amount of calculation is small as compared with a method using a noise codebook storing a plurality of noise vectors. In recent years, it has been used in various standard systems because it has the same sound quality as the conventional system.
[0007]
[Problems to be solved by the invention]
However, in the algebraic structure codebook, as the coding bit rate (coding rate) decreases, the deterioration of sound quality becomes noticeable. One of the reasons is lack of pulse position information. In other words, although the position information of the pulse is algebraically simplified in the algebraic structure codebook, the above advantage is obtained. However, at a low coding rate, a position candidate exists at a position where a pulse does not need to be raised, , The sound quality is degraded as well as the efficiency is low.
[0008]
Another reason that sound quality is degraded when an algebraic structure codebook is used is a shortage of pulses. When the number of pulses is insufficient, the noise "bubble wrap" becomes noticeable in the decoded speech. This is because the drive signal is generated from the pulse train, and as the number of pulses decreases, the presence / absence of the pulse becomes more audible. In order to improve the sound quality, it is necessary to reduce the bubble wrap.
[0009]
As described above, the conventional algebraic structure codebook has a simple structure and an advantage that the amount of calculation is small.On the other hand, when the coding rate becomes low, the position information and the number of pulses of the pulse train constituting the driving signal of the synthesis filter become low. There is a problem that the sound quality of the decoded voice is deteriorated due to the shortage.
[0010]
SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech encoding / decoding method capable of obtaining good sound quality even at a low encoding rate.
[0011]
[Means for Solving the Problems]
The present invention provides a step of generating information representing at least characteristics of a synthesis filter of an audio signal, and a signal for driving the synthesis filter, from a pulse position candidate that adaptively changes according to the characteristic of the audio signal. Generating a drive signal including a pulse train generated by arranging pulses at a selected predetermined number of pulse positions.
[0012]
According to the present invention, a driving signal including a pulse train generated by arranging pulses at a predetermined number of pulse positions selected from pulse position candidates that adaptively change according to the properties of an audio signal is input to a synthesis filter. To provide an audio decoding method for decoding an audio signal.
[0013]
In the speech encoding / decoding method according to the present invention, the drive signal for driving the synthesis filter arranges pulses at a predetermined number of pulse positions selected from pulse position candidates that change adaptively according to the properties of the speech signal. And the pulse train generated by the operation. More specifically, the pulse position candidates are arranged such that the greater the power of the audio signal, the more candidates there are.
[0014]
In addition, the drive signal is configured to include a pulse train generated by arranging pulses at all pulse position candidates that adaptively change according to the properties of the audio signal and optimizing the amplitude of each pulse by a predetermined means. You can also. In this case, more specifically, the pulse position candidates are arranged such that there are more candidates as the power of the audio signal increases.
[0015]
Further, the drive signal is a pulse train generated by arranging pulses at a predetermined number of pulse positions selected from first pulse position candidates that adaptively change according to the properties of the audio signal, or Using any one of the pulse trains generated by arranging pulses at a predetermined number of pulse positions selected from the second pulse position candidates consisting of part or all of the positions not used as the pulse position candidates It can also be generated. In this case, more specifically, the first pulse position candidates are arranged such that there are more candidates as the power of the audio signal is higher.
[0016]
When the drive signal includes a pitch vector and a noise vector, the noise vector is generated by arranging pulses at a predetermined number of pulse positions selected from pulse position candidates that change according to the shape of the pitch vector. You. In this case, more specifically, the pulse position candidates are arranged such that the greater the power of the pitch vector, the more candidates there are.
[0017]
Further, the noise vector is configured using a pulse train generated by arranging pulses at a predetermined number of pulse positions selected from position candidates set based on the position candidate density function determined from the shape of the pitch vector. You can also. In this case, more specifically, the pulse position candidates are arranged such that there are more candidates as the position candidate density function value increases, and the position candidate density function is the probability of the pitch vector power and the pulse being arranged. Is a function obtained in advance.
[0018]
Further, when a correction means such as a pitch cycle emphasis filter is used for the noise vector, a predetermined pitch selected from pulse position candidates that change in accordance with the shape of the inversely corrected pitch vector obtained by performing processing based on the inverse characteristic on the pitch vector. Is generated by arranging the pulses at the pulse positions of the number. In this case, more specifically, the pulse position candidates are arranged so that there are more candidates as the power of the inverse correction pitch vector is larger.
[0019]
By adaptively changing the pulse position candidates according to the properties such as the power distribution of the audio signal in this way, even when using an algebraic structure codebook in which the pulse positions and the number of pulses are reduced due to a lower coding rate. The coding efficiency is improved, and the coding rate can be reduced while maintaining the sound quality of the decoded speech. In addition, by using a pitch vector to generate a pulse position candidate, it is possible to adapt the pulse position candidate without requiring additional information.
[0020]
In another speech encoding / decoding method according to the present invention, when the drive signal includes a pitch vector and a noise vector, the method includes a pulse train shaped by pulse shaping means having characteristics determined based on the shape of the pitch vector. A drive signal is generated.
[0021]
With such a configuration, the pulse-like noise included in the decoded speech due to the decrease in the number of pulses is reduced, and the sound quality of the decoded speech is maintained even when the pulse position and the number of pulses are reduced by lowering the coding rate. It is possible to reduce the coding rate.
[0022]
Furthermore, in the speech encoding / decoding method according to the present invention, the speech signal is generated by arranging pulses at a predetermined number of pulse positions selected from pulse position candidates that adaptively change according to the properties of the speech signal. A drive signal including the generated pulse train may be generated, and the pulse train may be shaped by a pulse shaping unit having characteristics determined based on the shape of the pitch vector.
[0023]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 shows a speech encoding system to which the speech encoding method according to the first embodiment is applied. This speech coding system includes input terminals 101 and 106, an LPC analysis unit 110, an LPC quantization unit 111, an LPC synthesis unit 120, an auditory weighting unit 130, an adaptive codebook 141, and a pulse position candidate search unit. 142, an adaptive algebraic structure codebook 143, a code selector 150, a pitch period enhancer 160, gain multipliers 102 and 103, and adders 104 and 105.
[0024]
An input audio signal to be encoded is input to the input terminal 101 in units of a length corresponding to one frame, and the LPC analysis unit 110 performs linear prediction analysis in synchronization with the input audio signal, thereby corresponding to vocal tract characteristics. A linear prediction coefficient (LPC coefficient) is obtained. The LPC coefficient is quantized by an LPC quantization unit 111, and the quantized value is input to the LPC synthesis unit 120 as synthesis filter information indicating the characteristics of the LPC synthesis unit 120, and an index A indicating the quantization value is encoded. As a result, it is output to a multiplexing unit (not shown).
[0025]
The adaptive codebook 141 stores drive signals that have been input to the LPC synthesis unit 120 in the past. The drive signal input to the LPC synthesis unit 120 is a signal obtained by quantizing the prediction residual signal in the linear prediction analysis, and corresponds to a vocal cord signal including information of a sound pitch. The adaptive codebook 141 cuts out a waveform having a length corresponding to a pitch period from a past drive signal, and generates a pitch vector by repeating this. The pitch vector is usually obtained in units of subframes obtained by dividing a frame into several parts.
[0026]
The pulse position candidate search unit 142 calculates, based on the pitch vector obtained by the adaptive codebook 141, a position in the subframe at which the pulse position candidate is to be set, and calculates the result in the adaptive algebraic structure codebook 143. Output to
[0027]
The adaptive algebraic structure codebook 143 performs a predetermined operation so that distortion of an input speech signal from which the influence of the pitch vector is subtracted from the pulse position candidates input from the pulse position candidate search unit 142 is minimized under the auditory weight. Are searched for the pulse positions and the sign of the pulse positions.
[0028]
The pulse train output from the adaptive algebraic structure codebook 143 is cycled by a pitch unit by the pitch cycle emphasizing unit 160 as necessary. The pitch period emphasizing unit 160 receives the information L of the pitch period obtained by searching the adaptive codebook 143 from the input terminal 106, and gives the pulse train the periodicity of the pitch period.
[0029]
The pitch train output from the adaptive codebook 141 and the pulse train output from the adaptive algebraic structure codebook 143 and given a periodicity by the pitch period emphasizing unit 160 as necessary are converted into pitch vectors by the gain multiplying units 102 and 103. Are multiplied by a gain G0 for the noise vector and a gain G1 for the noise vector, respectively, added by the adding unit 104, and input to the LPC combining unit 120 as a drive signal. Note that, as the gains G0 and G1, usually, an optimum gain is selected from a gain codebook (not shown) storing a plurality of gains.
[0030]
The code selection unit 150 outputs an index B indicating the pitch vector selected in the search for the adaptive codebook 141, an index C indicating a pulse train selected in the search for the adaptive algebraic codebook 143, and a search for the gain codebook. An index G indicating the selected gains G0 and G1 is output. These indices B, C, G and the index A indicating the synthesis filter information, which is the quantization value of the LPC coefficient from the LPC quantization unit 111, are multiplexed by a multiplexing unit (not shown) and output as a bit stream.
[0031]
Next, the pulse position candidate search unit 142 and the adaptive algebraic structure codebook 143, which are characteristic parts of the present embodiment, will be described.
[0032]
In the present embodiment, even if the position where the pulse rises at the time of the low encoding rate is limited, the pulse is used to reduce only the encoding rate without deteriorating the sound quality as in the related art. The pulse position candidates are set for each sub-frame so that the position where the power of the drive signal is higher is allocated more position candidates by utilizing the property of standing concentrated on the position where the power is higher.
[0033]
Since the pitch vector is similar to the shape of an ideal drive signal, it is effective to set the pulse position candidate in the pulse position candidate search unit 142 based on the pitch vector obtained by searching the adaptive codebook 141. . Since the same pitch vector is required on the decoding side as on the encoding side, it is not necessary to generate extra additional information with the adaptation of the pulse position candidates.
[0034]
If the position candidates are assigned only to places where the power is large when adapting the pulse position candidates, the sound quality may be degraded due to the absence of the position candidates continuously in the section where the power is small. Various methods of adapting the pulse position candidates are conceivable. For example, by adopting the following method, adaptation with less sound quality deterioration is possible.
[0035]
With reference to the flowchart shown in FIG. 2, a description will be given of a processing procedure of pulse position candidate adaptation by the pulse position candidate search unit 142. FIG. 3 shows the input pitch vector waveform (F0), the power (F1) of this input pitch vector waveform, the smoothed power (F2), and the value obtained by integrating the smoothed power in the sample direction in each step of FIG. (F3) is shown corresponding to FIG.
[0036]
Similar processing can be performed by using other scales representing the waveform shape such as the absolute value of the amplitude value (the square root of the power) in addition to the power. In the present invention, these are collectively represented by power.
[0037]
First, the power (F1) is calculated for the input pitch vector (F0) of FIG. 3 (step S1), and then the power (F1) is smoothed to obtain a smoothed power (F2) (step S2). For power smoothing, for example, there is a method of taking a moving average by weighting in a window of several samples.
[0038]
Next, the power smoothed in step S2 is integrated in the sample direction (step S3). This situation is shown in (F3) of FIG. Specifically, assuming that the smoothed power of the n-th sample is p (n), the integrated value of the smoothed power p (n) is q (n), and the subframe length is L, the integrated value is q (n) is
q (n) = p (n) + q (n-1) + C (n = 0,..., L-1)
Is required. Here, C is a constant, and adjusts the degree of bias in the density of the pulse position candidates.
[0039]
Next, a pulse position candidate is calculated using the integrated value q (n) (step S4). In this case, the integral value is normalized such that the number of position candidates for which the integral value in the final sample is obtained is M. The position of the m-th candidate can be obtained as Sm by associating it with the integral value as shown in (F3) of FIG. By repeating the processing until m = 0,..., M−1, M position candidates can be obtained.
[0040]
FIG. 4 shows the relationship between the pulse position candidates thus obtained and the power of the pitch vector. The solid line indicates the power envelope of the pitch vector, and the arrows indicate pulse position candidates. As shown in the figure, the distribution of the pulse position candidates becomes dense where the power of the pitch vector is large, and becomes sparse as the power becomes small. As a result, where the power of the pitch vector that is important for sound quality is large, the pulse position can be selected more accurately. Further, even if the number of pulse position candidates decreases due to the lowering of the coding rate, high-quality encoding becomes possible by adaptively concentrating a small number of pulse position candidates in a place where the power of the pitch vector is large.
[0041]
Next, the position candidates thus obtained are distributed for each channel (step S5). Although there are various distribution methods, as shown in (F4) of FIG. 3, it is desirable that the position candidates are distributed so that each channel is alternated. In this way, the adaptive algebraic structure codebook 143 is obtained. In the search, the optimal position and code are selected for each pulse from each channel (Ch1, Ch2, Ch3) of the adaptive algebraic structure codebook 143, and a noise vector composed of three pulses is generated.
[0042]
When the sub-frame length is 80 samples, even if the pulse candidate positions are reduced to about 40 samples in all channels in total, the above method hardly causes auditory deterioration.
[0043]
In the algebraic structure codebook, the pulse amplitude is usually either +1 or -1, but a method using a pulse having amplitude information has also been proposed. Reference 4 (Chang Deyuan, "An 8 kb / s low complexity ACELP speech") codec, "1996 3rd International Conference on Signal Processing, pp. 671-4, 1996), with pulse amplitudes of 1.0, 0.5, 0, -0.5, -1.0. There is a method of selecting from among them. Also, reference 5 (K. Ozawa and T. Araseki, "Low Bit Rate Multi-pulse Speech Coder with Natural Speech Quality," IEEE Proc. ICASP. 86, p. The multi-pulse method, which is one of the above, is also composed of a pulse train having a drive signal having an amplitude. The present invention is also applicable to a case where a pulse as represented by these examples has an amplitude.
[0044]
Next, a speech decoding system corresponding to the speech encoding system of FIG. 1 will be described with reference to FIG.
[0045]
The parts having the same functions as those in FIG. 1 are denoted by the same reference numerals. The speech decoding system in FIG. 5 includes an LPC synthesizing unit 120, an LPC dequantizing unit 121, an adaptive codebook 141, a pulse position candidate The search unit 142, the adaptive algebraic structure codebook 143, the pitch period emphasis unit 160, the gain multiplication units 102 and 103, and the addition unit 104, and the coded stream transmitted from the speech coding system of FIG. Is entered.
[0046]
The input coded stream is input to a demultiplexing unit 121 (not shown), and the index A of the synthesis filter information and the index B indicating the pitch vector selected in the search for the adaptive codebook 141 by the demultiplexing unit 121. , An index C indicating a pulse train selected in the search for the adaptive algebraic structure codebook 143, an index G indicating the gains G0 and G1 selected in the search for the gain codebook, and an index L indicating the pitch period.
[0047]
The index A is decoded by the LPC inverse quantization unit 121 to obtain an LPC coefficient, which is synthesis filter information, and is input to the LPC synthesis unit 120. The indexes B and C are input to the adaptive codebook 141 and the adaptive algebraic structure codebook 143, respectively, and a pitch vector and a pulse train are output from these codebooks 141 and 143. In this case, the adaptive algebraic structure codebook 143 obtains the pulse position and code from the adaptive algebraic structure codebook 143 and the index B generated by the pulse position candidate search unit 142 based on the pitch vector input from the adaptive codebook 141. And outputs a pulse train. The pulse train output from the adaptive algebraic structure codebook 143 is given a periodicity of the pitch period L by the pitch period emphasizing unit 160 as necessary.
[0048]
The pitch train output from the adaptive codebook 141 and the pulse train output from the adaptive algebraic structure codebook 143 and given a periodicity by the pitch period emphasizing unit 160 as necessary are converted into pitch vectors by the gain multiplying units 102 and 103. Are multiplied by a gain G0 for the noise vector and a gain G1 for the noise vector, respectively, and then added by an adder 104 and input to the LPC synthesizer 120 as a drive signal, and the LPC synthesizer 120 outputs a reproduced audio signal. The gains G0 and G1 are selected from a gain codebook (not shown) according to the index G.
[0049]
As described above, according to the present embodiment, it is possible to reduce only the bit rate while maintaining the audio quality, and it is possible to realize audio encoding / decoding with high audio quality at a low encoding rate.
[0050]
FIG. 6 shows a speech encoding system according to the second embodiment of the present invention. This speech coding system removes the pulse position candidate search unit 142 and the adaptive algebraic structure codebook 143 from the configuration shown in FIG. 1 according to the first embodiment, and replaces the adaptive algebraic structure codebook 143 with general noise. A code book 144 is provided, and a pulse shaping filter analyzing unit 161 and a pulse shaping unit 162 are further added.
[0051]
Next, the processing procedure of the present embodiment is the same as that of the first embodiment up to the point of performing LPC analysis and LPC quantization of an input speech signal and then searching the adaptive codebook 141. In this example, the noise codebook 144 is configured by, for example, an algebraic structure codebook.
[0052]
The pulse shaping filter analysis unit 161 determines and outputs a filter coefficient of the pulse shaping unit 162 based on the pitch vector obtained by searching the adaptive codebook 141. Pulse shaping section 162 shapes the output of noise codebook 144 and outputs the result as a noise vector.
[0053]
Similarly to the first embodiment, the noise vector is periodicized using the pitch period emphasizing unit 160 as necessary, the gains G0 and G1 for the pitch vector and the noise vector are determined, and the index is output. Since the filter coefficient of the pulse shaping section 162 is obtained from the pitch vector, no new additional information is required.
[0054]
The feature of this embodiment lies in that the pulse shaping unit 162 is set based on the waveform of the pitch vector, and pulse shaping is performed on the pulse train output from the noise codebook 144 including the algebraic structure codebook. As described in the first embodiment, the pulse position and the number of pulses are reduced as the encoding rate is reduced, and the deterioration of the sound quality becomes conspicuous. When the number of pulses is reduced, the noise of “bubble wrap” becomes conspicuous in the decoded speech, but by using the pulse shaping unit 162 as in the present embodiment, the wrapping feeling is greatly reduced.
[0055]
Various methods can be used as a design method of the pulse shaping unit 162. As a first example, a method is conceivable that utilizes the property that when a drive signal for driving a synthesis filter is phase-equalized, it becomes a pulse-like signal. When an inverse filter for phase equalization is used, a pulse-like signal is input to obtain a drive signal-like waveform. A disadvantage of using a conventional pulse waveform is that phase information included in an ideal drive signal is lacking. This problem becomes more conspicuous as the number of pulses decreases. Therefore, by adding the phase information by the pulse shaping unit 162 as in this example, it is possible to generate a waveform closer to an ideal drive signal from the pulse waveform.
[0056]
In the first example, it is necessary to transmit the information of the filter coefficient of the phase equalization inverse filter, and the coding rate (bit rate) increases accordingly. Thus, as a second example of the pulse shaping unit 162, a method using a pitch vector as an approximation of the phase information is considered. Since the pitch vector is similar in shape to the drive signal in a sound section or the like, phase information can be extracted.
[0057]
As a specific method, a pulse shaping filter that determines a synchronization point such as a peak position of a pitch vector, extracts a waveform of several samples from the synchronization point, and uses the waveform as an impulse response can be used. The effect appears when the length of the extracted waveform is about two to three samples. It is also effective to apply a window to the sample taken out and attenuate it. Furthermore, since the same pitch vector can be obtained on the decoding side as on the encoding side, there is also an advantage that a new transmission bit is not required. At the time of searching noise codebook 144, pulse shaping section 162 is constant, so that the amount of calculation can be reduced by calculating the impulse response in advance together with LPC synthesis section 120.
FIG. 7 shows a speech decoding system corresponding to the speech encoding system of FIG. The parts having the same functions as those in FIG. 6 will be described with the same reference numerals. The speech decoding system in FIG. 7 includes an LPC synthesis unit 120, an LPC inverse quantization unit 121, an adaptive codebook 141, an algebraic structure code A noise codebook 144 composed of a book, a pulse shaping filter analyzing unit 161, a pulse shaping unit 162, a pitch period emphasizing unit 160, gain multiplying units 102 and 103, and an adding unit 104 are included. An encoded stream transmitted from the system is input.
[0059]
The input coded stream is input to a demultiplexing unit (not shown), and the index A of the synthesis filter information described above by the demultiplexing unit, the index B indicating the pitch vector selected in the search for the adaptive codebook 141, The index C representing the pulse train selected in the search for the noise codebook 144 and the index G indicating the gains G0 and G1 selected in the search for the gain codebook are separated and extracted. The pitch period L is calculated from the index B.
[0060]
The index A is decoded by the LPC inverse quantization unit 121 to become synthesis filter information, and is input to the LPC synthesis unit 120. The indexes B and C are input to the adaptive codebook 141 and the noise codebook 144, respectively, and a pitch vector and a pulse train are output from these codebooks 141 and 144.
[0061]
In this case, the pulse train output from the noise codebook 144 is processed by the pulse shaping unit 162 in which the coefficient is set by the pulse shaping filter analysis unit 161 based on the pitch vector obtained by the search of the adaptive codebook 141. The pitch period L is given a periodicity by the pitch period emphasizing unit 160 as necessary.
[0062]
The pulse train output from the adaptive codebook 141 and the pulse train output from the noise codebook 144 and passed through the pulse shaping unit 162 and the pitch period emphasizing unit 160 are processed by the gain multiplying units 102 and 103 with respect to the gain G0 for the pitch vector and the noise vector. After being multiplied by the gains G1, respectively, they are added by the adder 104, input to the LPC synthesizer 120 as a drive signal, and the LPC synthesizer 120 outputs a decoded speech signal synthesized. The gains G0 and G1 are selected from a gain codebook (not shown) according to the index G.
[0063]
As described above, according to the present embodiment, by using the pulse shaping unit 162, even when the algebraic structure codebook in which the number of pulses is reduced due to the lower coding rate is used as the noise codebook 144, the sound quality of the decoded speech is improved. It is possible to effectively reduce only the coding rate while maintaining it.
[0064]
FIG. 8 shows a speech coding system according to the third embodiment of the present invention. This speech coding system has a configuration obtained by adding the pulse shaping filter analysis unit 161 and the pulse shaping unit 162 described in the second embodiment to the configuration of the first embodiment.
[0065]
Next, the processing procedure of the present embodiment will be described. As in the first embodiment, first, LPC analysis and LPC quantization are performed, and after the search of the adaptive codebook 141 is completed, the pitch vector becomes the pulse position candidate search. It is passed to the section 142 and the pulse shaping filter analyzing section 161. The pulse position candidate search unit 142 obtains pulse position candidates using the method described in the first embodiment, and creates an adaptive algebraic structure codebook 143. In the pulse shaping filter analysis unit 161, the coefficients of the pulse shaping unit 162 are obtained as described in the second embodiment.
[0066]
In the search of the adaptive algebraic structure codebook 143, the output pulse train is shaped by the pulse shaper 162. In the actual search, the impulse responses of the pulse shaping unit 162 and the pitch period emphasizing unit 160 are combined with the LPC synthesizing unit 120, and the amount of calculation is reduced.
[0067]
FIG. 9 shows a speech decoding system corresponding to the speech encoding system of FIG. Since the operation of this speech decoding system is obvious from the operation of the speech decoding system described in the first and second embodiments, the same reference numerals are given to the same parts as in FIGS. Detailed description is omitted.
[0068]
Thus, in the present embodiment, the pulse position candidate searching unit 142 and the adaptive algebraic structure codebook 143 described in the first embodiment, and the pulse shaping filter analyzing unit 161 and the pulse shaping unit 162 described in the second embodiment. Are used at the same time, it is possible to maintain high sound quality even when a small number of pulses are set at limited position candidates, and it is possible to realize a speech coding method with high sound quality and a low coding rate.
[0069]
FIG. 10 shows a block diagram of a speech coding system according to the fourth embodiment of the present invention. In this speech coding system, the pulse position candidate search unit of the first embodiment is configured by a pitch vector smoothing unit 171, a position candidate density function calculation unit 172, and a position candidate calculation unit 173, except that It has the same configuration as the form.
[0070]
Next, the processing procedure of the present embodiment will be described. Similar to the first embodiment, after the LPC analysis and LPC quantization and the search of the adaptive codebook 141 are completed, the pitch vector is changed to the pulse position candidate search unit. 142 is passed to the pitch vector smoothing unit 171. The pitch vector smoothing unit 171 performs, for example, the processing of steps S1 and S2 of the flowchart of FIG. 2 on the pitch vector, obtains the power envelope of the pitch vector, and outputs this. The position candidate density function calculation unit 172 converts the power envelope into a position candidate density function and outputs it. The position candidate calculation unit 173 calculates a pulse position candidate using this position candidate density function instead of the power envelope, and creates an adaptive algebraic structure codebook 143 according to the obtained pulse position candidate. Subsequent processing is the same as in the first embodiment.
[0071]
The feature of the present embodiment lies in the processing method of the pulse position candidate search unit 142. In the first embodiment, the adaptation of the pulse position candidate is performed using the power envelope of the pitch vector as it is. In the present embodiment, the power envelope is converted into a position candidate density function, and then the adaptation is performed using the function. Is going. This will be described in detail with reference to FIG. FIG. 11A shows the power envelope of the pitch vector output from the pitch vector smoothing unit 171. The position candidate density function calculator 172 generates a position candidate density function (FIG. 11B) from the power envelope of the pitch vector (FIG. 11A). At this time, the conversion is performed using the function f indicating the correspondence between the value (x) of the power envelope and the value (f (x)) of the position candidate density function shown in FIG. As a method of creating the function f, for example, there is a method of statistically obtaining a number by processing a large number of learning voices.
Further, table data or the like can be used instead of the function.
[0072]
Since the pulse position candidate search unit 142 prepares the same one for the encoder and the decoder together with the function f for conversion, there is no need to send information on adaptation, and the There is no increase in bit rate.
[0073]
FIG. 12 shows the configuration of the speech decoding system of the present embodiment corresponding to the speech encoding system of FIG. Since the operation of the speech decoding system is obvious from the operation of the speech decoding system described in the first to third embodiments, a detailed description will be omitted.
[0074]
As described above, in the present embodiment, since the value of the power envelope of the pitch vector and the density of the pulse position candidate are converted using the function f, the processing procedure is slightly complicated as compared with the first embodiment, but is more accurate. Distribution of the position candidates can be realized. Further, the first embodiment can be considered as a case where x = f (x) in the present embodiment.
[0075]
FIG. 13 shows a block diagram of a speech coding system according to the fifth embodiment of the present invention. In this speech coding system, the pulse position candidate search unit of the first embodiment is configured by a pitch filter inverse operation unit 174, a smoothing unit 175, and a position candidate calculation unit 173. It has the same configuration.
[0076]
Next, the processing application of the present embodiment will be described. As in the first embodiment, after the LPC analysis and LPC quantization and the search of the adaptive codebook 141 are completed, the pitch vector becomes the pulse position candidate search. It is passed to the pitch filter inverse operation unit 174 of the unit 142. The pitch filter inverse operation unit 174 performs an operation representing an inverse characteristic of the pitch cycle emphasis unit 160. For example, the transfer function P (Z) of the pitch filter is
P (z) = 1−az ＾ (− L) (1)
When the transfer function Q (z) is obtained by the pitch filter inverse operation unit 174,
Q (Z) = 1 / (1-baz ＾ (− L)) (2)
And a method using a filter given by Here, a is a constant, b represents the degree of the inverse characteristic, and when b = 1, Q (z) is an inverse filter of P (z). The input pitch vector is output after being subjected to an inverse operation, and the power envelope is obtained by the smoothing unit 175 in the same manner as the pitch vector smoothing unit 171 of the fourth embodiment. The position candidate calculation unit 173 selects a pulse position candidate according to the power envelope, and creates an adaptive algebraic structure codebook 143. Subsequent processing is the same as in the first embodiment.
[0077]
A feature of the present embodiment is that a pitch vector in consideration of the influence of the pitch period emphasis unit 160 is used for adaptation of a pulse position candidate. The reason why the efficiency is improved by doing this will be described.
[0078]
The noise vector generated from the adaptive algebraic structure codebook is pitch-periodized by the pitch period emphasis unit 160. When the equation (1) is used for the periodization, the pulse near the head of the subframe is repeated many times within the subframe at the pitch cycle interval, whereas the number of repetitions decreases in the latter half of the pulse. By observing the actually obtained noise code vector, it can be confirmed that a pulse is likely to be generated at a position closer to the head when a strong pitch filter is used. This indicates that the pulse position is closely related not only to the pitch vector shape but also to the pitch filter. In the present embodiment, the use of the pitch filter inverse operation unit 174 realizes the adaptation of the pulse position candidate in consideration of the influence of the pitch period emphasis unit 160.
[0079]
By the way, in the third embodiment, it is possible to apply two types of filters, a pulse shaping filter and a pitch filter, to the noise vector. When the present embodiment is applied to such a case, it is ideal that a characteristic obtained by combining two filters is obtained, and an inverse characteristic of the characteristic is used for the pitch filter inverse operation unit. However, the effect can be obtained only by using only the characteristics of the pitch filter which has a large influence because the processing amount increases. Further, the order of the pitch filter inverse operation unit 174 and the smoothing unit 175 can be reversed.
[0080]
FIG. 14 shows the configuration of the speech decoding system of the present embodiment corresponding to the speech encoding system of FIG. Since the operation of the speech encoding system is obvious from the operation of the speech decoding system described in the first to fourth embodiments, detailed description will be omitted.
[0081]
FIG. 15 shows a block diagram of a speech coding system according to the sixth embodiment of the present invention. This speech coding system has the same configuration as that of the first embodiment, except that the adaptive algebraic structure codebook of the first embodiment is replaced by a noise vector generator 180 and an amplitude codebook 181.
[0082]
Next, the processing procedure of this embodiment will be described. First, after the LPC analysis and LPC quantization and the search of the adaptive codebook 141 are completed, the pitch vector is transmitted to the pulse position search unit 174 as in the first embodiment. Passed. The pulse position search unit 174 obtains a pulse position based on the power envelope of the pitch vector in the same manner as in the first embodiment, and outputs this to the noise vector generation unit. Here, the present embodiment is different from the previous embodiments in that the noise vector search unit generates all the pulses at the position obtained by the pulse position search unit 174. That is, in the embodiments described above, pulse position candidates are obtained, and the optimum pulse position is selected from the candidate in the adaptive algebraic structure codebook. On the other hand, in the present embodiment, all the pulse position candidates are used simultaneously. . Therefore, the process of selecting the pulse position becomes unnecessary. Instead, a process of selecting the amplitude of each pulse from the amplitude codebook 181 is added. As the output signal, information D representing the amplitude of the pulse is output instead of the information c representing the pulse position.
[0083]
A method of generating a noise vector will be described in detail with reference to FIG. FIG. 16A shows an amplitude pattern obtained from the amplitude codebook by an arrow. In this case, it is assumed that seven pulses are made. The waveforms in FIGS. 16B and 16C are the pitch vector power envelope obtained by the pulse position search unit 174 and the corresponding pulse positions (indicated by ○ in the figure). In FIG. 16 (b), there are two peaks in the power, so that seven pulse positions are dispersed in two places. In FIG. 16 (c), there is one peak in the center, so that the pulse position is focusing. FIGS. 16D and 16E are noise vectors in which a pulse having the amplitude of FIG. 16A is set at each pulse position. It can be seen that the shape of the drive signal also changes according to the pitch vector power envelope. As described above, since it is not necessary to transmit the power envelope information of the pitch vector, in the present embodiment, the shape of the noise vector can be approximated to the ideal noise vector shape without increasing the bit rate.
[0084]
In the present embodiment, as the bit rate increases, more pulse amplitude information D can be sent, and the quality is improved, but the degree of improvement is reduced. At a somewhat high bit rate, performance may be improved by including noise vectors with pulses at unselected positions as search candidates rather than increasing amplitude information. Specifically, the pulse position search unit 174 outputs a pattern (pulse pattern) of a different pulse position, and the noise vector generation unit searches for an amplitude for each pulse pattern. As the pulse pattern, in addition to the pulse pattern adapted to the above-described pitch vector, a pulse pattern generated from a pulse position not selected as the pulse pattern is prepared. For example, there is a method in which the remainder obtained by subtracting the sample position selected by the adaptation from all the sample positions of the subframe is used as a second pulse pattern to search for the amplitude of two types of pulse patterns. The number of bits assigned to the amplitude information may be different for each pulse pattern, and it is generally more efficient to allocate more bits to the pulse pattern using the adaptation. When a plurality of pulse patterns are used, it is necessary to transmit information indicating which pulse pattern was used in the information D, and the amplitude information is reduced accordingly, but only a single pulse pattern is searched. Better quality.
[0085]
FIG. 17 shows the configuration of the speech decoding system of the present embodiment corresponding to the speech encoding system of FIG. Since the operation of the speech decoding system is obvious from the operation of the speech decoding system described in the first to fifth embodiments, detailed description will be omitted.
[0086]
In the above embodiment, the speech encoding / decoding method has been described. However, the present invention can also be applied to a speech synthesis method. In that case, in the speech decoding system shown in FIGS. 5, 7, and 9, What is necessary is just to give each index based on the reproduced audio signal to be synthesized.
[0087]
【The invention's effect】
As described above, according to the present invention, high-quality speech encoding / decoding can be performed even when using an algebraic structure codebook in which the pulse position and the number of pulses are reduced by reducing the encoding rate.
[Brief description of the drawings]
FIG. 1 is a block diagram of a speech encoding system according to a first embodiment of the present invention.
FIG. 2 is a flowchart illustrating a pulse position candidate selection procedure according to the first embodiment;
FIG. 3 is a diagram showing a state of processing in each step of FIG. 2;
FIG. 4 is a diagram illustrating a relationship between a power envelope of a pitch vector and a pulse position candidate according to the first embodiment.
FIG. 5 is a block diagram of a speech decoding system according to the first embodiment;
FIG. 6 is a block diagram of a speech encoding system according to a second embodiment of the present invention.
FIG. 7 is a block diagram of a speech decoding system according to a second embodiment;
FIG. 8 is a block diagram of a speech encoding system according to a third embodiment of the present invention.
FIG. 9 is a block diagram of a speech decoding system according to a third embodiment.
FIG. 10 is a block diagram of a speech coding system according to a fourth embodiment of the present invention.
FIG. 11 is a diagram showing the relationship between the values of the pitch vector power envelope, the position candidate density function, the power envelope, and the position candidate density function.
FIG. 12 is a block diagram of a decoding system according to a fourth embodiment;
FIG. 13 is a block diagram of a speech encoding system according to a fifth embodiment of the present invention.
FIG. 14 is a block diagram of a decoding system according to a fifth embodiment;
FIG. 15 is a block diagram of a speech encoding system according to a sixth embodiment of the present invention.
FIG. 16 is a diagram for explaining a noise vector generation method.
FIG. 17 is a block diagram of a decoding system according to a sixth embodiment;
[Explanation of symbols]
101 ... Audio input terminal
102, 103 ... gain multiplication unit
104, 105 ... addition section
110 ... LPC analysis unit
111 LPC quantizer
120 ... LPC synthesis unit
130 ... auditory weighting unit
141 ... Adaptive codebook
142... Pulse position candidate search unit
143 ... Adaptive algebraic structure codebook
144: Noise codebook
150 ... code selection unit
160 pitch pitch emphasis unit
161: pulse shaping filter analyzer
162 ... Pulse shaping unit
171: pitch vector smoothing unit
172 position candidate density function calculator
173: position candidate calculation unit
174: pulse position search unit
180 ... Noise vector generation unit
181 ... amplitude codebook

Claims (8)

Generating synthesis filter information based on the input audio signal in frame units; generating a pitch vector from drive signals stored in the adaptive codebook for each subframe obtained by dividing the frame; Generating a pulse train by arranging pulses at a predetermined number of pulse positions selected from pulse position candidates arranged such that there are more candidates as large as possible, and a pitch vector of the adaptive codebook. A speech encoding method comprising: synthesizing the pulse train to generate a new drive signal; and generating a synthesized speech from the synthesis filter information and the new drive signal. 2. The speech encoding method according to claim 1, further comprising the step of giving a periodicity of a pitch period to said pulse train. The speech encoding method according to claim 1, wherein the driving signal generating step includes a step of multiplying the pitch vector and the pulse train by a gain. 4. The speech encoding method according to claim 3, further comprising: multiplexing an index indicating the synthesis filter information, an index indicating the pitch vector, an index indicating the pulse train, and an index indicating the gain to generate a bit stream. . The speech encoding method according to claim 1, further comprising a step of pulse-shaping the pulse train according to a filter coefficient determined based on the pitch vector. The pulse train generating step includes a step of obtaining a power envelope of the pitch vector, a step of converting the power envelope into a position candidate density function, and a step of calculating the pulse position candidate using the position candidate density function. The speech encoding method according to claim 1. Reproducing the synthesis filter information from the speech coded information; generating a pitch vector from the adaptive codebook based on the coded information; and Generating a pulse train by arranging pulses at a predetermined number of pulse positions selected from pulse position candidates arranged such that there is a new drive signal by synthesizing the pitch vector and the pulse train An audio decoding method including: generating; and generating a reproduced audio signal from the synthesis filter information and the new drive signal. 8. The speech decoding method according to claim 7, further comprising a step of pulse-shaping the pulse train according to a filter coefficient determined based on the pitch vector.

Images