JP2009104169A - Conversion device and conversion method of speech code string - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a conversion device of a speech code string, capable of converting a code string by a small arithmetic amount. <P>SOLUTION: In addition to a circuit for combining a decoding signal from the code string of a Code-Excited Linear Prediction (CELP) system at an input side, a circuit in which an LP coefficient and a pitch period respectively decoded by an LP coefficient decoding circuit 12 and a pitch component decoding circuit 13 are directly forwarded to each of the LP coefficient coding circuit 31 and a pitch component calculation circuit 40 at an output side, and supplied for code string conversion at the output side, is added. Thereby, LP analysis and selection of a pitch period candidate, which are conventionally performed for the decoding signal at the output side become unnecessary. When a band expansion processing is necessary at each of input and output sides, a circuit of band expansion conversion and pitch period candidate generation is added, and a coding circuit is provided instead of pitch component calculation. When a frame length of the input side is longer than that of the output side in the LP coefficient and the pitch period, interpolation is performed, and when it is shorter, average processing is performed. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a code string conversion device for converting a voice code string obtained by encoding one system into a voice code string decodable by the other system when performing voice communication between two types of voice encoding systems. In particular, the present invention relates to a speech code string conversion apparatus and a code string conversion method capable of converting into a speech code string with low distortion and a small amount of computation.

  Currently, there is a CELP (Code Excited Linear Prediction) method as a speech encoding method most frequently used in mobile phones and the like. References relating to the CELP system include “Code-Excited Linear Prediction: High Quality Speech at Very Low Bit Rates” (IEEE Proc. ICASSP-85, pp. 937-940, 1985) (hereinafter referred to as Non-Patent Document 1). There is.

  In a coding apparatus using the CELP method, a linear prediction (LP) coefficient representing a spectral envelope characteristic calculated by linear prediction (LP) analysis of an input speech signal, and an excitation signal for driving an LP synthesis filter composed of the LP coefficient, Encoding is performed separately. LP analysis and LP coefficient encoding are performed for each frame of a predetermined length. The encoding excitation signal is encoded for each subframe obtained by further dividing this frame into subframes.

  Here, the excitation signal is composed of a periodic component representing the pitch period of the input signal, the remaining residual component, and their gains. The periodic component representing the pitch period of the input signal is represented by an adaptive code vector stored in a code book that holds past excitation signals called an adaptive code book. The residual component is represented by a multi-pulse signal composed of a plurality of pulses called a sound source code vector or a signal designed in advance. The information of the sound source code vector is accumulated in the sound source code book.

  In a CELP decoding device, the decoded pitch period component and the excitation signal calculated from the residual signal are input to a synthesis filter composed of the decoded LP coefficients to obtain a synthesized speech signal.

  When communicating between two different CELP systems, as a conventional conversion device that converts a speech code string obtained by encoding of one system into a speech code string that can be decoded by the other system, decoding of one CELP system There is a conversion apparatus that obtains an output speech code string by encoding a speech signal decoded from a speech code string input by the apparatus using the other CELP method.

  Next, a conventional speech code string converting apparatus of this type will be described with reference to FIG. FIG. 11 is a block diagram illustrating a configuration example of a conversion apparatus that converts one CELP system A speech code string into another CELP system B speech code string.

  The illustrated conversion device includes an input terminal 10, a demultiplexer circuit 11, an LP coefficient decoding circuit 12, a pitch component decoding circuit 113, a residual component decoding circuit 14, and a speech synthesis circuit 15 for CELP A decoding processing. Prepare for. In addition, the frame circuit 21, the subframe circuit 22, the LP analysis circuit 130, the LP coefficient encoding circuit 31, the pitch period candidate selection circuit 132, the pitch component encoding circuit 41, the residual component encoding circuit 51, and the excitation signal synthesis circuit 52 The multiplexer circuit 53 and the output terminal 50 are provided for performing CELP system B encoding processing.

  The input terminal 10 inputs a CELP system A code string for each frame of the CELP system A and passes it to the demultiplexer circuit 11. The demultiplexer circuit 11 separates each code from the code string passed from the input terminal 10. The demultiplexer circuit 11 separates the code of the separated quantized LP coefficient to the LP coefficient decoding circuit 12, the code of the pitch period to the pitch component decoding circuit 113, and further the code of the residual component signal to residual component decoding Each is passed to the circuit 14.

  The LP coefficient decoding circuit 12 decodes the LP coefficient representing the spectral characteristics using the code passed from the demultiplexer circuit 11 and passes the decoded coefficient to the speech synthesis circuit 15.

  As an LP coefficient encoding and decoding method, there is a method of performing vector quantization after changing the LP coefficient into a line spectrum pair (LSP). In vector quantization, an encoder and a decoder have the same quantization vector table, and a code assigned to each vector is transmitted. The decoder outputs a vector corresponding to the passed code. For details of LSP vector quantization, see “Efficient Vector Quantization of LPC Parameters at 24 Bits / Frame” (IEEE Proc. ICASSP-91, pp. 661-664, 1991) (hereinafter referred to as Non-Patent Document 2). You can refer to it.

  The pitch component decoding circuit 113 decodes the pitch period L and the pitch gain ga from the code passed from the demultiplexer circuit 11. The pitch period L and the pitch gain ga are each scalar quantized, and a value corresponding to each passed code is searched from a previously designed quantization table to be a decoded value. The pitch component decoding circuit 113 accumulates the excitation signal passed from the speech synthesis circuit 15 up to samples for the past pitch period L, and extracts the accumulated excitation signal retroactively by the past pitch period L, thereby adapting the adaptive code vector Ca. Create Finally, the pitch component signal Ea (= ga · Ca) is calculated and passed to the speech synthesis circuit 15.

  The residual component decoding circuit 14 decodes the excitation code vector Cr and the excitation gain gr using the code passed from the demultiplexer circuit 11, calculates the residual component signal Er (= gr · Cr), and generates a speech synthesis circuit. Pass to 15. The sound source gain gr is scalar quantized, and a value corresponding to the passed code is searched from a previously designed quantization table to obtain a decoded value. As the sound source code vector Cr, a vector corresponding to the passed code is searched from a previously created sound source code book to be a decoded vector.

  The speech synthesis circuit 15 calculates the excitation signal vector Ex of the following Equation 1 using the pitch component signal Ea passed from the pitch component decoding circuit 113 and the residual component signal Er passed from the residual component decoding circuit 14. To the pitch component decoding circuit 113.

  Further, the speech synthesis circuit 15 is configured by the LP coefficient a (i) passed from the LP coefficient decoding circuit 12, and the excitation signal vector Ex previously calculated by the synthesis filter H (z) expressed by the following formula 2 is used. Filtering is performed to obtain a CELP system A decoded signal, and this decoded signal is passed to the frame circuit 21.

数式２で「ｐ」はＬＰ係数の次数である。

In Equation 2, “p” is the order of the LP coefficient.

  In the CELP system, in order to improve the audible sound quality, a filter that emphasizes a spectrum peak called a post filter is applied to the decoded signal. However, when encoding is performed again, this post filter is not applied to increase encoding distortion.

  The frame circuit 21 cuts out the decoded signal passed from the speech synthesis circuit 15 with the CELP system B frame length, and passes it to the LP analysis circuit 130, the pitch period candidate selection circuit 132, and the subframe circuit 22. The subframe circuit 22 divides the decoded signal passed from the frame circuit 21 into subframe lengths of CELP system B, and passes them to the pitch component encoding circuit 41.

  The LP analysis circuit 130 performs LP analysis on the decoded signal passed from the frame circuit 21 to obtain an LP coefficient. Next, the LP analysis circuit 130 passes the obtained LP coefficient to the LP coefficient encoding circuit 30 and the pitch period candidate selection circuit 132.

  The LP coefficient encoding circuit 31 vector-quantizes the LP coefficient passed from the LP analysis circuit 130 and passes the code to the multiplexer circuit 53. Reference can be made to the above Reference 2 for this quantization method. Further, the LP coefficient encoding circuit 31 passes the quantized LP coefficient to the pitch component encoding circuit 41 and the residual component encoding circuit 51.

  The pitch cycle candidate selection circuit 132 selects a pitch cycle candidate using the decoded signal passed from the frame circuit 21 and passes it to the pitch component encoding circuit 41. In the candidate selection, first, the decoded signal passed from the frame circuit 21 is filtered by the weighting filter W (z) expressed by the following Equation 3 composed of the LP coefficient a (i) passed from the LP analysis circuit 130. . In Equation 3, “β” and “γ” are coefficients for adjusting the load for performing auditory sound quality improvement, and take values that satisfy “0 <γ <β ≦ 1”.

  Next, the pitch period candidate selection circuit 132 calculates the autocorrelation function of the weighted decoded signal in the range of the correlation lag “20 to 147”, and calculates the correlation lag that maximizes the autocorrelation and its neighboring values to the pitch. Candidate for period.

  The pitch component encoding circuit 41 encodes the pitch period component of the decoded signal vector Sd having a subframe length passed from the subframe circuit 22 for each subframe, and passes the code to the multiplexer circuit 53. The pitch component encoding circuit 41 first creates an adaptive code vector by cutting back the previously decoded excitation signal passed from the residual component encoding circuit 51 by a time L by a subframe length. Next, the pitch component encoding circuit 41 filters the adaptive code vector by the above-described equation 2 to calculate a decoded signal Sa (L) having only the pitch component. Further, the pitch component encoding circuit 41 loads the decoded signal vector Sd and the pitch period component vector Sa (L) using Equation 3 above, and loads the load decoded signal vector Sdw and the load pitch period component vector Saw (L). Get.

  The pitch component encoding circuit 41 performs the above-described operation on the pitch cycle component for each of the pitch cycle candidates passed from the pitch cycle candidate selection circuit 132, and performs the load decoding signal vector Sdw and the load pitch cycle component vector Saw (L ) To determine the optimum pitch period Lo that minimizes the square distance Da.

  The square distance Da is obtained by the following formula 4 using the optimum pitch gain ga (L) calculated for each pitch period L. Further, the optimum pitch gain ga (L) is obtained by the following formula 5. In the following description, the symbol ‖x‖ means the norm of the vector x, and the symbol <x, y> means the inner product of the vector x and the vector y.

  The pitch component encoding circuit 41 finally passes a code obtained by scalar quantization of the optimum pitch period Lo and the corresponding pitch gain ga (Lo) to the multiplexer circuit 53.

  Further, the pitch component encoding circuit 41 obtains a vector obtained by integrating the optimum pitch gain gaq (Lo) quantized to the load pitch period component vector Saw (Lo) from the load decoded signal vector Sdw. The residual signal vector Sdw ′ is passed to the residual component encoding circuit 51. Further, the pitch component encoding circuit 41 synthesizes a pitch component excitation signal E′a obtained by integrating the quantized optimum pitch gain gaq (Lo) to the adaptive code vector Ca (Lo) corresponding to the optimum pitch period Lo. Pass to circuit 52.

  The residual component encoding circuit 51 encodes the residual signal vector Sdw ′, which is the residual component of the decoded signal vector Sd passed from the pitch component encoding circuit 41, for each subframe, and the code is sent to the multiplexer 53. hand over.

  That is, the residual component encoding circuit 51 first extracts the k-th excitation code vector Cr (k) from the excitation codebook designed and stored in advance. Next, the residual component encoding circuit 51 filters the excitation code vector by the above-described Equation 2 and calculates a decoded signal Sr (k) having only the residual component. Further, the residual component encoding circuit 51 loads each of the decoded signal vector Sd and the residual component vector Sr (k) using Equation 3 above, and the weighted decoded signal vector Sdw and the weighted residual component vector Srw (k). And get. The residual component encoding circuit 51 performs the above-described operation relating to the residual component on all the excitation code vectors stored in the excitation codebook, and the residual signal vector passed from the pitch component encoding circuit 41. The code ko of the excitation code vector that minimizes the square distance Dr between Sdw ′ and the load residual component vector Srw (k) is determined.

  The square distance Dr is obtained by the following formula 6 using the optimum sound source gain gr (k) calculated for each delay. The optimum sound source gain gr (k) is obtained by the following formula 7.

  Finally, the residual component encoding circuit 51 scalar quantizes the optimum excitation gain gr (ko), and passes the code and the code ko of the excitation code vector to the multiplexer circuit 53. Also, the residual component encoding circuit 51 supplies the residual component excitation signal E′r obtained by integrating the optimal excitation gain grq (ko) quantized to the selected excitation code vector Cr (ko) to the excitation signal combining circuit 52. hand over.

  The excitation signal synthesis circuit 52 adds the pitch component excitation signal E′a passed from the pitch component encoding circuit 41 and the residual component excitation signal E′r passed from the residual component encoding circuit 51. The excitation signal Ex ′ is calculated by the following formula 8 and passed to the pitch component encoding circuit 41.

  The multiplexer circuit 53 connects the codes obtained by the respective encodings passed from the LP coefficient encoding circuit 31, the pitch component encoding circuit 41, and the residual component encoding circuit 51 in a predetermined order, and connects the code string. Generate and pass to output terminal 50. The output terminal 50 outputs the code string passed from the multiplexer circuit 53.

“Code-Excited Linear Prediction: High Quality Speech at Very Low Bit Rates” IEEE Proc. 1985, ICASSP-85, pp. 937-940 “Efficient Vector Quantization of LPC Parameters at 24 Bits / Frame” IEEE Proc. 1991, ICASSP-91, pp. 661-664)

  However, the above-described conventional speech code string conversion apparatus has a problem that the code conversion processing amount is large and an increase in size cannot be avoided.

  The reason for this is that when a decoded signal obtained by synthesizing the encoded code string from the demultiplexer circuit through the decoding circuit is encoded by the CELP method B on the output side through the frame circuit in the CELP method A on the input side. This is because the code string conversion is performed on all parameters via the synthesized decoded signal.

  An object of the present invention is to solve such problems and to provide a speech code string conversion apparatus and method that can convert a speech code string with a small amount of computation without increasing distortion.

  In order to solve the above problems, a speech code string converter according to the present invention inputs a first code string including a pitch period from an input terminal as an input side, and converts the first code string including a pitch period into a second code string including a pitch period. Output from the output terminal on the output side, and the first means includes a pitch included in the first code string for each subframe which is a time unit for encoding the pitch period of the second code string. A pitch component decoding circuit on the first code string side and a pitch component calculation circuit on the second code string side, each having a period as a pitch period included in the second code string, are provided.

  The second means calculates, for each subframe, which is a time unit for encoding the pitch period of the second code string, from the pitch period in the subframe of the first code string and the pitch period in the past subframe. A circuit for interpolating or averaging the pitch period to be calculated, and a pitch component calculating circuit for setting the calculated pitch period to a pitch period included in the second code string are provided.

  The third means obtains a pitch period included in the first code string and at least a plurality of pitch period candidates in the vicinity thereof for each subframe which is a time unit for encoding the pitch period of the second code string. A pitch cycle candidate generation circuit to be generated, and a pitch component encoding circuit that uses any one of the generated candidates as a pitch cycle included in the second code string are provided.

  For each subframe that is a time unit for encoding the pitch period of the second code string, the fourth means calculates the pitch period of the subframe of the first code string and the pitch period of the past subframe. A pitch period interpolation or averaging circuit for calculating a pitch period, a pitch period candidate generation circuit for generating the calculated pitch period and at least a plurality of pitch periods in the vicinity thereof as a pitch period candidate, and a generated candidate And a pitch component encoding circuit having a pitch period included in the second code string as one of them.

  According to a fifth means, the pitch component encoding circuit in the third or fourth means is an audio signal decoded from the first code string and an audio signal decoded from the second code string for each subframe. The pitch period included in the second code string is selected so as to minimize the distance between

  In the sixth means, the pitch component encoding circuit in the third or fourth means is decoded from the excitation signal decoded from the first code string and the second code string for each subframe. The pitch period included in the second code string is selected so as to minimize the distance from the excitation signal.

  In a seventh aspect, the spectral characteristic included in the first code string is changed to the spectral characteristic included in the second code string for each frame which is a time unit for encoding the spectral characteristic of the second code string. The LP coefficient decoding circuit is provided on the first code string side, and the LP coefficient coding circuit is provided on the second code string side.

  The eighth means obtains a spectral characteristic from the spectral characteristic of the frame of the first code string and the spectral characteristic of the past frame for each frame which is a time unit for encoding the spectral characteristic of the second code string. It is characterized by comprising a circuit for interpolation or averaging of LP coefficients to be calculated, and an LP coefficient encoding circuit that uses the calculated spectral characteristics as spectral characteristics included in the second code string.

  The ninth means includes, for each frame of the second code string, a band extension conversion circuit that converts the band extension intensity of the spectrum characteristic included in the first code string, and converts the spectrum characteristic obtained by the conversion into the second And an LP coefficient encoding circuit having spectral characteristics included in the code string.

  In addition, the tenth means, for each frame that is a time unit for encoding the spectrum characteristic of the second code string, calculates the spectrum from the spectrum characteristic of the frame of the first code string and the spectrum characteristic of the past frame. The second code string includes an LP coefficient interpolation or averaging circuit for calculating the characteristics, a band extension conversion circuit for converting the band extension intensity of the calculated spectrum characteristics, and the spectrum characteristics obtained by the conversion. An LP coefficient encoding circuit having spectral characteristics is provided.

  As described above, when the first code string is converted into the second code string, the present invention uses the LP coefficient decoded by the first code string as the LP analysis result in the second code string. Yes. As a result, the LP analysis process can be omitted in the second code string process. Also, the pitch period decoded by the first code string or a pitch period such as the vicinity thereof is used as a pitch period candidate in the second code string. As a result, the pitch cycle candidate selection process can be omitted in the second code string process.

  That is, according to the present invention, the LP coefficient and pitch period decoded from the input CELP code sequence are directly used on the output side, and the decoded signal obtained by decoding the input code sequence is used. Therefore, it is possible to eliminate the need for LP analysis and selection of pitch period candidates that were previously performed on the decoded signal on the input side. The effect of becoming.

It is a functional block diagram which shows the 1st Embodiment by this invention. It is a functional block diagram which shows the 2nd Embodiment by this invention. It is a functional block diagram which shows the 3rd Embodiment by this invention. It is a figure explaining the interpolation process of the LP coefficient in this invention. It is a figure explaining the interpolation process of the pitch period in this invention. It is a functional block diagram which shows the 4th Embodiment by this invention. It is a functional block diagram which shows the 5th Embodiment by this invention. It is a figure explaining the averaging process of LP coefficient in the present invention. It is a figure explaining the averaging process of the pitch period in this invention. It is a functional block diagram which shows the 6th Embodiment by this invention. It is a functional block diagram which shows an example of the past.

  Next, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a functional block diagram showing a first embodiment of the present invention. This form is a case where the frame length and the subframe length of CELP system A and CELP system B match.

  The illustrated speech code string conversion apparatus includes an input terminal 10, a demultiplexer circuit 11, an LP coefficient decoding circuit 12, a pitch component decoding circuit 13, a residual component decoding circuit 14, and a speech synthesis circuit 15 in a CELP system A. Prepare for decryption. The frame circuit 21, the subframe circuit 22, the LP coefficient encoding circuit 31, the pitch component calculation circuit 40, the residual component encoding circuit 51, the excitation signal synthesis circuit 52, the multiplexer circuit 53, and the output terminal 50 are CELP-type. It is provided to perform the B encoding process.

  The difference from FIG. 11 referred to as the conventional converter is that the LP analysis circuit 130 and the pitch period candidate selection circuit 132 are deleted, the pitch component decoding circuit 113 is replaced with the pitch component decoding circuit 13, and the pitch component encoding circuit 41 is replaced with The pitch component calculation circuit 40 is changed.

  In this code string converter, the input terminal 10 inputs a CELP system A code string and passes it to the demultiplexer circuit 11. The demultiplexer circuit 11 separates the code string passed from the input terminal 10, passes the code of the quantized LP coefficient to the LP coefficient decoding circuit 12, passes the code of the pitch component to the pitch component decoding circuit 13, and The sign of the residual component signal is passed to the residual component decoding circuit 14.

  The LP coefficient decoding circuit 12 decodes the LP coefficient representing the spectral characteristics using the code passed from the demultiplexer circuit 11, and passes the decoded coefficient to the speech synthesis circuit 15 and the LP coefficient encoding circuit 31. The pitch component decoding circuit 13 decodes the pitch period L and the pitch gain ga from the code passed from the demultiplexer circuit 11. The pitch component decoding circuit 13 is different from the pitch component decoding circuit 113 of FIG. 11 only in passing the pitch period L to the pitch component calculation circuit 40. In addition, the excitation signal passed from the speech synthesis circuit 15 is accumulated up to the samples of the past pitch period L, and the accumulated excitation signal is cut back by the pitch period L in the past to create the adaptive code vector Ca. Finally, the pitch component signal Ea (= ga · Ca) is calculated and passed to the speech synthesis circuit 15.

  The residual component decoding circuit 14 decodes the excitation code vector Cr and the excitation gain gr using the code passed from the demultiplexer 11 to calculate the residual component signal Er (= gr · Cr), and the speech synthesis circuit 15 To pass. The speech synthesizing circuit 15 calculates the excitation signal vector Ex of Equation 1 using the pitch component signal Ea passed from the pitch component decoding circuit 13 and the residual component signal Er passed from the residual component decoding circuit 14. To the pitch component decoding circuit 13. Further, the speech synthesis circuit 15 filters the excitation signal vector Ex with the synthesis filter H (z) according to the above equation 2 composed of the LP coefficient a (i) passed from the LP coefficient decoding circuit 15 to obtain a decoded signal vector. Sd is obtained and passed to the frame circuit 21.

  The frame circuit 21 cuts out the decoded signal passed from the speech synthesis circuit 15 by the frame length of CELP system B and passes it to the subframe circuit 22. The subframe circuit 22 divides the decoded signal passed from the frame circuit 21 into subframe lengths of CELP system B, and passes them to the pitch component calculation circuit 40.

  The LP coefficient encoding circuit 31 quantizes the LP coefficient passed from the LP coefficient decoding circuit 12 and passes the code to the multiplexer circuit 53. Further, the LP coefficient encoding circuit 31 passes the quantized LP coefficient to the pitch component calculation circuit 40 and the residual component encoding circuit 51.

  The pitch component calculation circuit 40 generates an adaptive code vector by cutting back the previously decoded excitation signal passed from the excitation signal synthesis circuit 52 by the time L by a subframe length. Next, the pitch component calculation circuit 40 filters the adaptive code vector with the above-described Equation 2 to calculate the decoded signal Sa (L) having only the pitch component. Further, the pitch component calculation circuit 40 loads the decoded signal vector Sd and the pitch cycle component vector Sa (L) using the above Equation 3, and obtains the load decoded signal vector Sdw and the load pitch cycle component vector Saw (L). . The pitch component calculation circuit 40 calculates the pitch gain ga (L) by using the above formula 5 using these values. Finally, the pitch component calculation circuit 40 passes the code obtained by scalar quantization of the pitch period L and the pitch gain ga (L) to the multiplexer circuit 53. Further, the pitch component signal E′a calculated by the product of the quantized pitch gain gaq (L) and the adaptive code vector Caq (L) is passed to the excitation signal synthesis circuit 52.

  The residual component encoding circuit 51 encodes the residual component of the decoded signal vector Sd passed from the pitch component calculation circuit 40 for each subframe, and passes the code to the multiplexer 53.

  First, the residual component encoding circuit 51 extracts the k-th excitation code vector Cr (k) from the excitation code book designed and stored in advance. Next, the residual component encoding circuit 51 filters the excitation code vector with the above-described Equation 2 to calculate a decoded signal Sr (k) having only the residual component. Further, the residual component encoding circuit 51 loads the decoded signal vector Sd and the residual component vector Sr (k) using Equation 3 to load the decoded decoded signal vector Sdw and the weighted residual component vector Srw (k). ) And get.

  The residual component encoding circuit 51 performs the above-described operation related to the residual component on all the excitation code vectors stored in the excitation codebook, and the residual signal vector Sdw passed from the pitch component calculation circuit 40. The square distance Dr between 'and the load residual component vector Srw (k) is calculated using Equation 6 above, and the code ko of the minimum excitation code vector is determined.

  Finally, the residual component encoding circuit 51 scalar quantizes the optimum excitation gain gr (ko), and passes the code and the code ko of the excitation code vector to the multiplexer circuit 53. Also, the residual component encoding circuit 51 supplies the residual component excitation signal E′r obtained by integrating the optimal excitation gain grq (ko) quantized to the selected excitation code vector Cr (ko) to the excitation signal combining circuit 52. hand over.

  The excitation signal synthesis circuit 52 adds the pitch component excitation signal E′a passed from the pitch component calculation circuit 40 and the residual component excitation signal E′r passed from the residual component encoding circuit 51. 8 calculates the excitation signal Ex ′ and passes it to the pitch component calculation circuit 40.

  The multiplexer circuit 53 predetermines the LP coefficient, pitch period, pitch gain, excitation codebook, and excitation gain code passed from the LP coefficient encoding circuit 31, the pitch component calculation circuit 40, and the residual component encoding circuit 51. A code string is generated by connecting to the order and passed to the output terminal 50. The output terminal 50 outputs the code string passed from the multiplexer circuit 53.

  Next, a second embodiment of the present invention will be described with reference to FIG.

  This form adds a band extension conversion process for correcting that the spectrum band extension process differs between CELP system A and CELP system B, and a pitch period candidate generation process for generating pitch period candidates. .

  2 differs from FIG. 1 in that a band extension conversion circuit 30 and a pitch period candidate generation circuit 32 are added, and the pitch component encoding circuit 41 described in FIG. 11 is used instead of the pitch component calculation circuit 40. It is that you are. The band extension conversion circuit 30 is located between the LP coefficient decoding circuit 12 and the LP coefficient encoding circuit 31. The pitch period candidate generation circuit 32 is located between the pitch component decoding circuit 13 and the pitch component encoding circuit 41.

  In FIG. 2, the same constituent elements as those in FIG. Therefore, the band extension conversion circuit 30 and the pitch period candidate generation circuit 32 related to these processes will be described next.

  In the band expansion process, when calculating the LP coefficient a (i) from the autocorrelation function r (i) of the input signal so that a sharp peak does not occur in the spectral characteristics, the window function w (i) such as an exponential window is calculated. In this process, “w (i) · r (i)” is obtained by integrating the autocorrelation function r (i). Since this window function w (i) differs depending on the encoding method, the deterioration due to the conversion can be reduced by correcting this in the code string conversion. The pitch period candidate generation process is a process of selecting a pitch period decoded by CELP system A as it is from CELP system B without using it as it is in CELP system B. Compared with the case where the pitch period is used as it is, a calculation amount for determining the pitch period is required, but deterioration due to conversion can be reduced.

  The band extension conversion circuit 30 calculates the impulse response of the LP filter composed of the LP coefficients passed from the LP coefficient decoding circuit 12, and uses the CELP system A band extension coefficient wa (i) as the autocorrelation function of the impulse response. And the band expansion coefficient wb (i) of CELP system B is integrated. Next, the band extension conversion circuit 30 calculates the LP coefficient from the autocorrelation function by the Levinson-Durbin method or the like, and passes it to the LP coefficient encoding circuit 31.

  The pitch cycle candidate generation circuit 32 passes the pitch cycle L and the neighboring pitch cycle passed from the pitch component decoding circuit 13 to the pitch component encoding circuit 41 as pitch cycle candidates. In order to suppress deterioration in sound quality due to code string conversion, the passed pitch period may include an integer multiple of the pitch period L, a value of a fraction of an integer, or a value in the vicinity thereof as a pitch period candidate.

  The pitch component encoding circuit 41 receives the pitch period candidate from the pitch period candidate generation circuit 32 and performs the same operation as described in the conventional method. At this time, in order to reduce the amount of calculation, speech synthesis is performed using the optimum pitch gain G′a (L) calculated for each delay in order to omit the filtering according to Equation 2 and the load according to Equation 3. It is also possible to determine the optimal pitch period Lo that minimizes the square distance D′ a between the excitation signal Ex calculated by the circuit 15 and the adaptive code vector Ca (L).

  The square distance D'a is obtained using the following equation 9, and the optimum pitch gain G'a (L) is obtained using the following equation 10.

  Next, a third embodiment of the present invention will be described with reference to FIG.

  In this embodiment, the frame length Na and the subframe length Nsa of CELP system A are longer than the frame length Nb and the subframe length Nsb of CELP system B, respectively. This embodiment is different from the second embodiment in that it has a process of adjusting the difference between the frame length and the subframe length.

  3 differs from FIG. 2 in that an LP coefficient interpolation circuit 60 and a pitch period interpolation circuit 70 related to these processes are added. The LP coefficient interpolation circuit 60 is located between the LP coefficient decoding circuit 12 and the band extension conversion circuit 30. The pitch period interpolation circuit 70 is located between the pitch component decoding circuit 13 and the pitch period candidate generation circuit 32.

  In FIG. 3, the same components as those in FIG. The added LP coefficient interpolation circuit 60 and pitch period interpolation circuit 70 will be described next.

  To make the description concrete, the frame length Na of CELP system A is 20 ms, the subframe length Nsa is 10 ms, the frame length Nb of CELP system B is 10 ms, and the subframe length Nsb is 5 ms. Suppose that The LP coefficient is calculated by the LP analysis window with the final subframe of each frame as the center.

  The LP coefficient interpolation circuit 60 uses the LP coefficient passed every 20 ms of the frame length Na from the LP coefficient decoding circuit 12 and the LP coefficient passed in the past frame, every 10 ms of the frame length Nb used in the CELP system B. The LP coefficient is calculated and passed to the band extension conversion circuit 30.

  FIG. 4 shows the relationship between the LP coefficients of CELP system A and CELP system B. The X mark shown in the figure is the center of the LP analysis window described above, and is the center when the LP coefficient is interpolated. The frame number is indicated by “k” in CELP system A and “t” in CELP system B, respectively. The arrows indicate which LP coefficients of CELP scheme A are used to calculate the LP coefficients of CELP scheme A.

  The LP coefficient representing the spectral characteristics of the CELP system A frame is passed from the LP coefficient decoding circuit 12 every 20 ms, whereas the CELP system B requires an LP coefficient every 10 ms. For this reason, when “i” is “1, 2,..., Or p” as indicated by an arrow shown in FIG. 4, the LP coefficient ab (CELP method B at frame numbers “t−1” and “t” ( t-1, i) and LP coefficient ab (t, i) are converted to LP coefficient aa (k, i) of the corresponding frame in CELP system A, and LP coefficient aa (k -J, i) is used according to the following equations 11 and 12. In the calculation, a load function w (j) that defines an interpolation method is used. In consideration of the positional relationship of the X mark in the example shown in FIG. 4, in the case of the LP coefficient ab (t−1, i) in Expression 11, “w (0) = 5/8, w (1) = 3 / 8 "and" M = 2 "apply. In the case of the LP coefficient ab (t, i) in Expression 12, “w (0) = 1” and “M = 1” are applied.

  The pitch period interpolation circuit 70 determines the subframe length Nsb used in the CELP system B from the pitch period passed every 10 ms of the subframe length Nas from the pitch component decoding circuit 13 and the pitch period passed in the past subframe. A pitch period every 5 ms is calculated and passed to the pitch period candidate generation circuit 32.

  FIG. 5 shows the relationship between the pitch periods of CELP system A and CELP system B. The frame number is indicated by “k” in CELP system A and “t” in CELP system B, respectively. The arrows indicate which pitch period of CELP system B is used to calculate the pitch period of CELP system A.

  From the pitch component decoding circuit 13, the pitch period of the subframe of CELP system A is passed every 10 ms. However, since the CELP system B requires a pitch period every 5 ms, as shown by the arrow in FIG. The pitch period L1b (t) and pitch period L2b (t) of CELP system B in the first and second subframes with the number “t” are the pitch period L1a (k) and pitch period L2a of the corresponding frame in CELP system A. (K) and the pitch period L1a (k−j) and the pitch period L2a (k−j) in the frame that is traced back by j frames in the past, respectively, are calculated by the pitch period Lsb (t) of Equation 13 below. In the calculation, a load function u (j) that defines an interpolation method is used.

  Further, in consideration of the positional relationship between the CELP subframes in the example shown in FIG. 5, when the pitch period Lsb (t) in Expression 13 is the pitch period L1b (t), “u (0) = 3/4”. , U (1) = 1/4 ”and“ M = 2 ”apply. Further, when the pitch period Lsb (t) is the pitch period L2b (t), “u (0) = 1” and “M = 1” are applied.

  Next, a fourth embodiment of the present invention will be described with reference to FIG.

  This form is a case where the CELP system A frame length Na and the subframe length Nsa are longer than the CELP system B frame length Nb and the subframe length Nsb, respectively, as in the third embodiment.

A band extension conversion process for correcting that the spectrum band extension process differs between CELP system A and CELP system B and a pitch period candidate generation process for generating pitch period candidates are added.

  6 differs from FIG. 1 in that an LP coefficient interpolation circuit 60 and a pitch period interpolation circuit 70 are added, and the band extension conversion circuit 30 and the pitch period candidate generation circuit 32 are deleted from FIG. Instead of the pitch component encoding circuit 41, the pitch component calculation circuit 40 described with reference to FIG. 1 is used. Therefore, the LP coefficient interpolation circuit 60 is located between the LP coefficient decoding circuit 12 and the LP coefficient encoding circuit 31. The pitch period interpolation circuit 70 is located between the pitch component decoding circuit 13 and the pitch component calculation circuit 40.

  In FIG. 6, the same components as those in FIG. The LP coefficient interpolation circuit 60 and the pitch period interpolation circuit 70 are added to FIG. 1, but are functionally the same as those described with reference to FIGS.

  That is, the LP coefficient interpolation circuit 60 interpolates the LP coefficient passed from the LP coefficient decoding circuit 12 and passes it to the LP coefficient encoding circuit 31. The pitch period interpolation circuit 70 interpolates the pitch period passed from the pitch component decoding circuit 13 and passes it to the pitch component calculation circuit 40.

  Next, a fifth embodiment of the present invention will be described with reference to FIG.

  In this embodiment, the frame length Na and the subframe length Nsa of CELP system A are shorter than the frame length Nb and the subframe length Nsb of CELP system B, respectively. This embodiment differs from the second embodiment in that it has a process for adjusting the difference between the frame length and the subframe length, and the difference adjustment processing method is different from that in the third embodiment. Yes.

  That is, FIG. 7 differs from FIG. 3 in that the LP coefficient interpolation circuit 60 and the pitch period interpolation circuit 70 in FIG. 3 related to these processes are respectively replaced by the LP coefficient averaging circuit 61 and the pitch period averaging circuit 71. It has been replaced. Therefore, the LP coefficient averaging circuit 61 is located between the LP coefficient decoding circuit 12 and the band extension conversion circuit 30. The pitch period averaging circuit 71 is located between the pitch component decoding circuit 13 and the pitch period candidate generating circuit 32.

  In FIG. 7, the same components as those in FIG. 3 are given the same reference numerals and description thereof is omitted. Therefore, the replaced LP coefficient averaging circuit 61 and pitch period averaging circuit 71 will be described next.

  In order to make the description concrete here, the frame length Na of CELP system A is 10 ms, the subframe length Nsa is 5 ms, the frame length Nb of CELP system B is 20 ms, and the subframe length Nsb is 10 ms. Suppose that The LP coefficient is calculated by the LP analysis window with the final subframe of each frame as the center.

  The LP coefficient averaging circuit 61 uses the LP coefficient passed from the LP coefficient decoding circuit 12 every 10 ms of the frame length Na and the LP coefficient passed in the past frame, and every 20 ms of the frame length Nb used in the CELP system B. The LP coefficient is calculated and passed to the band extension conversion circuit 30.

  FIG. 8 shows the relationship between the LP coefficients of CELP system A and CELP system B. The X mark shown in the figure is the center of the LP analysis window described above, and is the center when the LP coefficients are averaged. The frame number is indicated by “k” in CELP system A and “t” in CELP system B, respectively. The arrows indicate which LP coefficients of CELP scheme A are used to calculate the LP coefficients of CELP scheme A.

  The LP coefficient decoding circuit 12 passes the LP coefficient representing the spectrum characteristics of the CELP system A frame every 10 ms, but the CELP system B requires an LP coefficient every 20 ms. For this reason, when “i” is “1, 2,..., Or p” as indicated by an arrow shown in FIG. 8, the LP coefficient ab (t, i) of the CELP system B at the frame number “t” is The calculation is performed according to Equation 12 using the LP coefficient aa (k, i) of the frame corresponding to CELP system A and the LP coefficient aa (k−i, i) of the frame that is traced back by j frames in the past. In the calculation, a load function w (j) that defines an averaging method is used. In consideration of the positional relationship of the X mark in the example shown in FIG. 8, in the case of the LP coefficient ab (t, i) in Expression 12, “w (0) = 3/4, w (1) = 1/4. And “M = 2” apply.

  The pitch period averaging circuit 71 uses the pitch period passed every 10 ms of the subframe length Nas from the pitch component decoding circuit 13 and the pitch period passed in the past subframe, and the subframe length Nsb used in the CELP system B. The pitch period is calculated every 5 ms and passed to the pitch period candidate generation circuit 32.

  FIG. 9 shows the relationship between the pitch periods of CELP system A and CELP system B. The frame number is indicated by “k” in CELP system A and “t” in CELP system B, respectively. The arrows indicate which pitch period of CELP system B is used to calculate the pitch period of CELP system A.

  From the pitch component decoding circuit 13, the pitch period of the subframe of CELP system A is passed every 5 ms, but CELP system B requires a pitch period every 10 ms. Therefore, as indicated by the arrows in FIG. 9, CELP method A corresponds to pitch period L1b (t) and pitch period L2b (t) of CELP method B in the first and second subframes of frame number “t”. Using the pitch period L1a (k) and the pitch period L2a (k) of the frame to be played, and the pitch period L1a (kj) and the pitch period L2a (kj) in the frame that is traced back by j frames in the past It is calculated by 13 pitch periods Lsb (t). In the calculation, a load function u (j) that defines an averaging method is used. Further, in consideration of the positional relationship between the CELP subframes in the example shown in FIG. 9, when the pitch period Lsb (t) in Expression 13 is the pitch period L1b (t), “u (0) = 1/2 , U (1) = 1/2 ”and“ M = 2 ”apply. Similarly, in the case of the pitch period L2b (t), “u (0) = 0, u (1) = 0, u (2) = 1/2, u (3) = 1/2” and “M = 4 "applies.

  Next, a sixth embodiment of the present invention will be described with reference to FIG.

  This form is a case where the CELP system A frame length Na and the subframe length Nsa are shorter than the CELP system B frame length Nb and the subframe length Nsb, respectively, as in the fifth embodiment. This embodiment is different from the second embodiment in that it has a process for adjusting the difference between the frame length and the subframe length, and the difference adjustment processing method is different from that in the fifth embodiment. Yes.

  10 is different from FIG. 1 in that an LP coefficient averaging circuit 61 and a pitch period averaging circuit 71 are added. The difference from FIG. 7 is that the band extension conversion circuit 30 and pitch period candidate generation are performed. The circuit 32 is deleted, and the pitch component calculation circuit 40 described with reference to FIG. 1 is used instead of the pitch component encoding circuit 41. Therefore, the LP coefficient averaging circuit 61 is located between the LP coefficient decoding circuit 12 and the LP coefficient encoding circuit 31. The pitch period averaging circuit 71 is located between the pitch component decoding circuit 13 and the pitch component calculation circuit 40.

  In FIG. 10, the same constituent elements as those in FIG. Further, the LP coefficient averaging circuit 61 and the pitch period averaging circuit 71 are added to FIG. 1, but are functionally the same as those described with reference to FIGS.

  That is, the LP coefficient averaging circuit 61 averages the LP coefficients passed from the LP coefficient decoding circuit 12 and passes them to the LP coefficient encoding circuit 31 in the same manner as described in the fifth embodiment. The pitch period averaging circuit 71 averages the pitch periods passed from the pitch component decoding circuit 13 and passes them to the pitch component calculation circuit 40.

  As described above, according to the present invention, the LP coefficient and pitch period decoded from the input CELP code sequence are directly used on the output side, and the decoding obtained by decoding the input code sequence Since code conversion is performed without using a signal, it is possible to eliminate the need for LP analysis and selection of pitch period candidates that were conventionally performed on a decoded signal on the input side. The effect that conversion becomes possible is acquired.

  INDUSTRIAL APPLICABILITY The present invention can be widely used in a CELP decoding apparatus that obtains a synthesized speech signal by inputting a decoded pitch period component and an excitation signal calculated from a residual signal into a synthesis filter composed of decoded LP coefficients. .

DESCRIPTION OF SYMBOLS 10 Input terminal 11 Demultiplexer circuit 12 LP coefficient decoding circuit 13 Pitch component decoding circuit 14 Residual component decoding circuit 15 Speech synthesis circuit 21 Frame circuit 22 Subframe circuit 30 Band extension conversion circuit 31 LP coefficient encoding circuit 32 Pitch period candidate generation Circuit 40 Pitch component calculation circuit 41 Pitch component encoding circuit 50 Output terminal 51 Residual component encoding circuit 52 Excitation signal synthesis circuit 53 Multiplexer circuit 60 LP coefficient interpolation circuit 61 LP coefficient averaging circuit 70 Pitch period interpolation circuit 71 Pitch period average Circuit

Claims (4)

  In a code string converter for converting a first code string including a spectrum characteristic into a second code string including a spectrum characteristic, a spectrum included in the first code string for each frame of the second code string A code string conversion device comprising: a circuit that converts a spectral characteristic obtained by converting a band extension strength of a characteristic into a spectral characteristic included in the second code string.   In the code string conversion device for converting the first code string including the spectral characteristics into the second code string including the spectral characteristics, for each frame which is a time unit for encoding the spectral characteristics of the second code string, A spectrum characteristic obtained by converting the band extension intensity of the spectrum characteristic calculated from the spectrum characteristic of the frame of the first code string and the spectrum characteristic of the past frame is included in the spectrum included in the second code string. A code string conversion device comprising a circuit having characteristics.   In a code string conversion method for converting a first code string including a spectrum characteristic into a second code string including a spectrum characteristic, a spectrum included in the first code string for each frame of the second code string A code string conversion method comprising: converting a band extension strength of a characteristic; and converting a spectral characteristic encoded after the conversion into a spectral characteristic included in the second code string.   In a code string conversion method for converting a first code string including a spectral characteristic into a second code string including a spectral characteristic, for each frame which is a time unit for encoding the spectral characteristic of the second code string, Calculating a spectral characteristic from a spectral characteristic in the frame of the first code string and a spectral characteristic in the past frame; converting a band extension intensity of the calculated spectral characteristic; And a step of converting the spectral characteristics included in the second code string.

Images