JP3453116B2 - Audio encoding method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding method and apparatus for digitizing a speech signal and coding a speech coding parameter representing the characteristic thereof at a predetermined time interval. A digital mobile phone or digital voice storage that transmits or stores voice coding parameters, restores voice coding parameters from the transmission destination or storage destination when needed, and synthesizes a voice signal from the restored voice coding parameters to transmit voice. It is suitable for use in devices and the like.

[0002]

2. Description of the Related Art Since a digitized voice signal can be subjected to various digital signal processing such as data compression, error processing, and multiplexing, it is not limited to fixed-line telephones and mobile telephones, but can be applied to multimedia systems using voice. Widely adopted. To digitize analog audio signals,
Generally, sampling is performed at a sampling frequency that is twice or more the input voice frequency band, and quantization is required at a quantization step that cannot be discriminated by the ear. Therefore, a wide transmission frequency bandwidth is required as compared with an analog signal. Therefore, the audio signal once digitized is compressed by various encoding and modulation methods according to the required audio quality.

As a method for obtaining a high compression rate of voice data, an analysis-synthesis type voice encoding system which positively uses the characteristics of voice and a voice encoding parameter obtained therefrom are efficiently quantized. A method is being considered. For example, MBE (Multi-
Band Excitation) method or IMBE (Improved Mu)
The lti-Band Excitation) method is a type of this analysis-synthesis type speech coding method, in which speech is transmitted at a predetermined time interval (20 msec).
Divide into segments to form a frame, and for each frame, the voice pitch (or the voice fundamental frequency as its inverse), the voice harmonics spectrum amplitude sequence obtained from the voice frequency spectrum of the frame, and the frequency spectrum are set to appropriate frequencies. Voiced / unvoiced information (Voiced / Unvoiced information or V / UV information) for each frequency band divided into regions is used as a voice encoding parameter, and for each frame, the voice pitch is 8-bit uniform quantization, and V for each band is used. / UV information v
[k] (k is the band number) is a binary value expressed as a binary number of 0/1, and K bit quantization (K: maximum number of bands is 12).
(Variable length of bits), the speech harmonics amplitude sequence is obtained by two-dimensionally transforming the inter-frame prediction difference value and quantizing its DCT (discrete cosine transform) coefficient by 75-K bits to obtain a speech coding rate of 4.15 kbps.

FIG. 7 is a diagram showing the configuration of a general voice coding transmission device. The sampled / quantized audio digital signal input from the audio input terminal 301 is divided into segments at predetermined time intervals by the audio encoding parameter extraction unit 302 to form a frame, and the audio encoding is performed for each frame. Extract the parameters. The speech coding parameters to be extracted differ depending on the speech coding method, and for example, the above M
In the BE method, it is a voice pitch, a voice harmonics spectrum amplitude sequence, and V / UV information of each frequency band. The parameter encoding unit 303 effectively encodes the extracted speech encoding parameter to reduce the code amount, and the transmitting unit 304
Is sent to the transmission line 305 via. With respect to the signal received by the receiving unit 306, a voice decoding parameter is restored by a parameter decoding unit 307, and a voice synthesizing unit 308 creates a synthesized voice by an operation reverse to that of the voice encoding parameter extracting unit 302, and outputs it from a voice output terminal 309. Outputs a digital audio signal.

FIG. 8 is a block diagram of the voice coding parameter extraction unit 302 in the case of the MBE method. The digital input voice signal is input from the input terminal 301 to the fundamental frequency estimation unit 401, where the fundamental frequency of the voice is estimated. The estimated value of the fundamental frequency is calculated as the reciprocal value of the time when the time-delayed autocorrelation function becomes maximum. The frequency spectrum calculation unit 402 frequency-analyzes the finite-length voice signal cut out from the frame using a window function such as a Hamming window to obtain a voice frequency spectrum. The fundamental frequency correction unit 403 sets Ab-S (Analysis) under the minimum error condition between the estimated speech fundamental frequency, the spectrum synthesized by the window function, and the speech frequency spectrum.
-by-Synthesis) method is used to simultaneously obtain the modified speech fundamental frequency ωo and the harmonic spectrum amplitude sequence. The voiced strength calculation unit 404 divides the frequency band into a plurality of frequency bands k (k = 1,2, ..., K) based on the modified speech fundamental frequency ωo, and synthesizes the synthesized spectrum for each frequency band. And the error of the voice frequency spectrum is calculated, and by the threshold judgment
V / UV information v [k] is output. Spectrum envelope calculation unit 40
5 is V / UV information v [k], which is Abs in voiced band.
The root mean square value of the frequency spectrum in each harmonics frequency band for each harmonics spectrum amplitude obtained by the method and the unvoiced band is the spectrum envelope absolute value | A (ω) |
Output as.

FIG. 9 is a block diagram of the parameter coding unit 303 in the case of the IMBE method.
The fundamental frequency quantization unit 502 uniformly quantizes the voice fundamental frequency ωo input to the input terminal 501 into 8 bits in a predetermined quantization range and quantization step, and outputs the quantized value B0 to the output terminal 503. To do. The V / UV information v [k] input to the input terminal 504 is the V / UV information quantizer 505.
If, for example, the number K of frequency bands is 12, 12 0's
Alternatively, it is output to the output terminal 506 as a binary 12-bit value B1 represented by the information of 1. The spectral envelope | A (ω) | input to the input terminal 507 is a discrete harmonics spectrum amplitude sequence | A (ωi) |, (i = 1,2,3, ..., N, N: number of harmonics. ). First, after the logarithmic transformation is performed by the logarithmic transformation unit 508, the difference (which will be referred to as “prediction difference value sequence”) from the predicted harmonics spectrum amplitude sequence 521 predicted from the harmonics spectrum amplitude sequence of the previous frame is calculated by the subtractor 509. And passed to the block conversion unit 510. The block transformation unit 510 sequentially classifies the prediction difference value sequence into 6 types according to the order of harmonics to form two-dimensional data, passes it to the next DCT (discrete cosine transformation) unit 511, calculates DCT coefficients, and passes it to the quantization unit 512. , DC
The quantized data B2 of the prediction difference value sequence is output to the output terminal 513 by the combination of the uniform quantization method and the vector quantization method selected in advance according to the order of the T coefficient. Quantization restoration unit 514, inverse DCT unit 515 and block restoration unit 516
Restores the quantized prediction difference value sequence and adds it to the adder 517.
And add it to the predicted harmonics spectrum amplitude column 521,
The quantized spectral envelope value of the input spectral envelope of the current frame is restored. The quantized spectrum envelope value is delayed by one frame in the frame delay unit 518, and based on the newly input voice fundamental frequency ωo of the next frame and the fundamental frequency of the previous frame, the spectrum envelope prediction unit 519 uses the spectrum envelope of the next frame. The value is predicted, the predicted value is guided to the subtractor 509, and the spectrum envelope for the next frame is quantized.

The principle of this MBE method is based on DW Griffin
and JS Lim, "Multi-band Excitation Vocoder", I
EEE Transactions on Acoustics, speech, and signal
processing, vol.36, No.8, August 1988, pp1223-1235. In addition, the configuration method of the encoder is USP-5491722 (Methods for
speech transmssion, Feb. 13, 1996). As described above, as a method of digitizing voice to realize low bit rate voice encoding, an analysis-synthesis type voice encoding method is proposed in which voice encoding parameters based on a voice synthesis model are extracted and encoded. , Partly put to practical use.

[0008]

Although the analysis-synthesis type speech coding system described above is effective for low bit rate speech coding, this method is based on a certain speech analysis / synthesis model. Since voice analysis / synthesis is performed only with the voice synthesis parameter, synthetic voice quality is likely to occur depending on the configuration of the encoding system. As a method of improving this point, a method of reducing the change of the voice parameter in the voice frame by setting the voice frame update period to be short and improving the sound quality while being of the analysis and synthesis type can be considered. Regarding the effect of improving the coded voice quality when the frame update cycle is set short, M. et al., “Voice coding system for commercial mobile communication”, IEICE 2000 National Convention, D14-2, p171 , Mar. 2000.

However, shortening the update cycle of the voice frame is contrary to the high compression rate (low bit rate) of voice. For example, in the above-mentioned report of Mitsuru et al., A speech frame is divided into two and the speech coding parameters are taken out as a sub-frame structure. When the speech coding parameters are coded by the conventional method, a double bit rate is required. Becomes Therefore, there is a demand for a parameter coding method in which the number of parameter coding bits does not become extremely large even when the voice segment length is set to be short. As described above, in order to improve the voice quality of the analysis-synthesis type speech encoding method aiming at a low bit rate, there is a problem of an efficient encoding method of the analysis-synthesis type speech encoding parameter, and particularly, the frame update period. There is a problem with measures to prevent an increase in the encoding bit rate when the value is shortened.

Therefore, an object of the present invention is to provide a speech coding method and apparatus of the analysis-synthesis type which can significantly reduce the coding bit rate. Further, in the analysis-synthesis type speech encoding method and apparatus, the speech-synthesis type which improves the encoded speech quality by shortening the frame updating period of speech encoding and prevents the increase in the number of encoded bits is provided. It is an object of the present invention to provide a voice encoding method and device.

[0011]

In order to solve the above-mentioned problems, a speech coding method of the present invention acquires a speech coding parameter from a speech signal which is digitized and divided into frames of a predetermined time length, and encodes it. A voice encoding method for
A plurality of voice pitches as the voice coding parameters are calculated using a frame to be quantized by any one of the quantization methods by selecting a differential quantization method and a uniform quantization method, and quantized voice pitches of preceding and following frames. Of any of the interpolated voice pitch candidates of, and encoding with a combination with a frame to be quantized by the index number, voiced / voice as the voice encoding parameter
The step of encoding the unvoiced information by the index number of the one having the shortest distance from the limited number of representative voiced / unvoiced information, and the harmonic spectrum amplitude sequence as the speech encoding parameter are calculated by the linear prediction model. Frames quantized by linear prediction coefficients (or line spectrum pairs derived from them) and gain, and multiple interpolated linear prediction coefficients (or line spectra) calculated using linear prediction coefficients (or line spectrum pairs) of preceding and following frames A pair) a step of selecting any one of a plurality of combinations of a candidate and a plurality of interpolation gain candidates calculated by using gains of preceding and following frames, and encoding by a combination of frames to be quantized by the index number.

Further, in the selection of the differential quantization method and the uniform quantization method in the quantization of the voice pitch, the differential quantization method is selected when the differential quantization error is less than a threshold value, and the quantum quantization method is selected when the differential quantization error is more than that threshold value. This is a process of selecting a quantization method with less quantization error, and the selection of the interpolated voice pitch candidate is performed by interpolation that is closest to the voice pitch of the current frame among a plurality of interpolated voice pitch candidates calculated from voice pitches of preceding and following frames. This is a process of selecting a voice pitch candidate. Further, the quantization of the voice pitch is performed using a logarithmic pitch obtained by logarithmically converting the voice pitch. Furthermore, the difference quantization of the voice pitch is performed by setting a larger quantization step as the difference value increases.

Furthermore, the limited number of representative voiced / unvoiced information items are deleted from a large number of pre-acquired voiced / unvoiced information items in the order of low occurrence frequency, and the removed voiced / unvoiced information items are deleted. It is assumed that it was created by reducing the frequency of occurrence to a desired limited number of voiced / unvoiced information by allocating and integrating the frequency of occurrence to adjacent voiced / unvoiced information according to the distance and size to it. There is. Furthermore, the voiced / unvoiced information distance is represented by 0 or 1 for voiced / unvoiced voices in each voice spectrum band, and values of 0 or 1 in each voice spectrum band are converted into binary bits in the order of frequency of the voice spectrum band. It is said that they are fitted and represented by the difference between their binary values. Furthermore, the voiced / unvoiced information is composed of a combination of a frame for transmitting voiced / unvoiced information and a frame for not transmitting it. Furthermore, in the frame in which the voiced / unvoiced information is omitted, the voiced / unvoiced information possessed by the frame having the larger energy or the average amplitude value of the frames before and after the frame is used for decoding. Furthermore, when the harmonics spectrum amplitude sequence is modeled by a linear prediction model, the harmonics spectrum amplitude value of the 0th order (DC component) is corrected, and then the linear prediction modeling of the harmonics spectrum amplitude is performed. . Furthermore, the plurality of interpolation gain candidates include a gain obtained by equally or unequally dividing the gains of the preceding and following frames, a gain having a higher gain and a gain having a lower gain in the preceding and following frames. .

Furthermore, the speech coding apparatus of the present invention comprises:
A voice encoding device for acquiring and encoding a voice encoding parameter from a voice signal which is digitized and divided into frames of a predetermined time length, wherein a voice pitch as the voice encoding parameter is uniform with a differential quantization method. A frame that is quantized by one of the quantization methods by selecting the quantization method,
Means for selecting one of a plurality of interpolated voice pitch candidates calculated using the quantized voice pitch of the preceding and following frames, and encoding by a combination of frames quantized by the index number, and the voice encoding parameter Means for encoding the voiced / unvoiced information as a voice by the index number of the closest distance from the limited number of representative voiced / unvoiced information, and the harmonics spectrum amplitude sequence as the voice encoding parameter, A frame that is quantized by a linear prediction coefficient (or a line spectrum pair derived from it) and a gain by a prediction model, and a plurality of interpolated linear prediction coefficients (using linear prediction coefficients (or line spectrum pairs) of preceding and following frames ( Or line spectrum pair) Candidates and gains of previous and next frames Select one of the combinations of the plurality of interpolated gain candidate, by its index number is intended to include means for encoding by the combination of a frame to be quantized.

As described above, in the present invention, with respect to the coding of the voice pitch, the logarithmically converted pitch is
Switching between the differential quantization method and the uniform quantization method, and selecting the one with less error from the input speech pitch for quantization,
By selecting the number of the closest interpolation point candidate from multiple interpolation points of the inter-frame voice pitch and quantizing with that selection number, by switching by using frame repetition, the number of quantization bits of the voice pitch can be changed. is decreasing. For voiced / unvoiced information (V / UV information) coding,
V / UV information and its frequency of occurrence are acquired in advance for many audio frames, a fixed number of representative V / UV information is selected from among them, and V / UV of each frame is selected from the representative V / UV information. It takes the means of encoding with the number (index) of the representative V / UV information that is most similar to the information. In addition, insert a frame that does not transmit V / UV information as appropriate, and decode the frame that does not have V / UV information transmitted. I am trying to use V / UV information. Furthermore, regarding the encoding of the harmonics spectrum amplitude sequence, the spectrum amplitude sequence is modeled by a high-order all-pole model, that is, an autoregressive linear prediction model (AR model), and its linear prediction coefficient (L
A means for quantizing the gain (PC coefficient) and the gain, or the LSP (line spectrum pair) obtained by modifying the LPC coefficient and the gain is used. Also, the LPC coefficient of the frame or L
For the quantization of the SP and the gain, a means for selecting the number of the nearest interpolation point candidate from the quantized LPC coefficients or a plurality of interpolation points between the frames of the LSP and the gain and quantizing with the selected number is also used. This reduces the number of quantization bits in the harmonic spectrum amplitude sequence.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the speech coding method and apparatus of the present invention will be described by taking as an example the case of being applied to the MBE or IMBE speech coding method which is the analysis-synthesis type speech coding method. This speech coding apparatus corresponds to the parameter coding unit 303 shown in FIG. 7, and the speech coding parameter extracted by the speech coding parameter extraction unit 302, that is, the speech pitch (or its reciprocal). Voice fundamental frequency ωo), voice harmonics spectrum amplitude sequence, and voiced / unvoiced information (V / UV information) of each frequency band are efficiently encoded.

FIG. 1 is a block diagram showing an example of the configuration of a speech coding apparatus to which the speech coding method of the present invention is applied.
For example, the voice pitch (or the voice fundamental frequency ωo) obtained by the voice coding parameter extraction unit shown in FIG. 8 is input to the input terminal 101, the voice pitch is logarithmically converted by the logarithmic conversion unit 102, and the logarithmic voice pitch P [n] (n is a frame number) is obtained. Logarithmic voice pitch is described in the literature (Thomas Eriksson
and Hong-Goo Kang, "Pitch Quantization in low Bit-
Rate Speech Coding ", ICASSP '99, pp489-492,1999)
It is known that the human detection limit value with respect to the change amount of the logarithmic pitch is not so much influenced by the value of the logarithmic pitch, as described in. Therefore, the quantization step width can be made uniform, which is a convenient conversion.

The logarithmic pitch P [n] is alternately switched by the input switching unit 103 for each frame (or each subframe when subframed), and the two output terminals 1
It is output to either 04 or 116. When it is output to 104, it is guided to the uniform quantization unit 105 and the subtraction unit 112. The uniform quantizing unit 105 quantizes uniformly at a certain quantizing step, and the quantized logarithmic pitch P1 ′ [n] is input to the pitch comparing unit 108. On the other hand, the subtraction unit 112 obtains the differential logarithmic pitch from the input logarithmic pitch and the quantized logarithmic pitch of the previous frame received from the delay unit 111, and inputs it to the differential quantization unit 113. The delay unit 111 is
It is for passing the quantized logarithmic pitch of the immediately preceding frame to the current frame. The differential quantization unit 113 performs differential quantization at a constant differential quantization step, or at a non-uniform quantization step set so that the differential quantization step expands as the differential input amplitude increases with zero input as symmetry.
The addition unit 114 adds the quantized logarithmic pitch of the previous frame as a reference, and the logarithmic pitch P2 ′ [n] obtained by differential quantization is 10
Output to 7.

In the pitch comparison unit 108, P1 '[n] and P2' [n]
Are compared, the quantized logarithmic pitch having a smaller error from the logarithmic pitch P [n] before quantization is selected, and the quantized logarithmic pitch P ′ [n] of this frame is output to the output terminal 109. To the output terminal 110, the index output from the quantizer selected by the pitch comparison unit 108 out of the uniform quantization index N1 and the differential quantization index N2 is output as a pitch code. By arranging the indexes of N1 and N2 so that the numbers do not overlap, it is possible to know which quantization method is selected from the output index numbers.

The other output 116 of the input switching unit 103
The logarithmic pitch P [n] appearing in the above is a plurality of interpolation pitch candidates created by the interpolation pitch candidate creating unit 117 using the quantized log pitches of the current frame and the previous frame obtained from the input / output terminal of the delay unit 111, The interpolation point index (selection number) N3 that is compared by the interpolation point comparison unit 119 and gives the pitch closest to the logarithmic pitch of the output 116 is set as 1 as the pitch interpolation code.
It is output to 20.

FIG. 2 shows the interpolation pitch candidate creation unit 11 of FIG.
It is a figure explaining the operation of 7. The example shown in FIG. 2 is an example in which the number of interpolation point candidates is 4, and quantization is performed with 2 bits according to the selection number. Quantize logarithmic pitch of the front frame of the current frame is P '[n + 1], Quantize logarithmic pitch of the rear frame is P' [n-1], and divide them evenly with straight lines. Is indicated by a cross. Of the four interpolation pitch candidates, the interpolation quantization pitch P ′ that is closest to the input logarithmic pitch P [n].
[n] is selected, and 2 is selected as an index giving this interpolation quantization pitch in the example of FIG. P [n] is P '[n
Is the pitch of the frame between +1] and P '[n-1], eg
If the frame is divided into two subframes, P [n] is the first subframe of the current frame, P '[n + 1] is the quantization pitch of the second subframe of the current frame, and P' [ n-
1] corresponds to the quantization pitch of the second subframe of the previous frame.

In the arrangement of interpolation pitch candidates in FIG. 2, P '
[n + 1] and P '[n-1] are not included in the interpolation pitch candidate, but the interpolation pitch candidate should be set to include P' [n + 1] and P '[n-1] at both ends. You can also In that case, there are two interpolation pitch candidates other than P '[n + 1] and P' [n-1]. When the positions of the interpolation pitch candidates are set excluding both ends as in the example of FIG. 2, two points excluding both ends can be selected even with 1 bit. Therefore, the interpolation point arrangement of FIG. Bit or 2
It can be said to be effective when there are few bits.

Returning to FIG. 1 again, voiced / unvoiced information (V / UV
Information) is input from the input terminal 131, and the frame thinning unit 133 thins out the V / UV information. Eg 2
V / UV information is output only once for each frame.
It is input to the unvoiced comparison unit 134. On the receiving side (decoding side), in the frame in which the voiced / unvoiced determination information is omitted, the voiced / unvoiced information of the frame having the largest frame energy or amplitude average value is used among the frames before and after the frame. Decrypt.

The representative voiced / unvoiced information codebook 132 is a method in which a limited number of occurrences are selected and stored according to the method of the present invention, which will be described later, from V / UV information acquired in advance from many voice frames. Currently entered V / UV
Voiced the representative V / UV information value b1 'at the closest distance to the information value b1
It is selected by the unvoiced comparison unit 134 and the index of the representative V / UV information is output as the voiced / unvoiced code 135. The V / UV information value b1 (or b1 ') is the V / UV information value v [k], k = 1,2, ... for each frequency band obtained by dividing the voice frequency spectrum at intervals of, for example, three times the voice fundamental frequency. .., K (v [k] is 0 or 1) is represented by a binary value assigned to each bit of the binary number.

[Equation 1] The distance of the V / UV information value is represented by the absolute value of the difference between b1 and b1 '. In addition, the representative V / UV information is set independently for each band number determined by the audio fundamental frequency.

The representative V / UV information can be selected and created by the following method. That is, a large number of V / UV information values are obtained in advance from a large number of audio frames, and the number of bands K is classified. A set of V / UV information values classified for each number of bands {b1 _i }
Then, the frequency of occurrence of each b1 _i is totaled. For example, 5 bands
In the case of, the value of b1 takes an integer value from 0 to 31,
The frequency of occurrence is calculated for each b1 value. From this, for example, when quantizing V / UV information with 2 bits, it is necessary to select four types of representative V / UV information from this. For this selection, a method of selecting four from the highest occurrence frequency may be considered, but depending on the case, the adjacent representative V /
The UV information value may be selected and may not be suitable as the representative V / UV information value. In particular, when the number of bands K is large, there is a high possibility that representative V / UV information values having a high occurrence frequency will exist adjacently.

Therefore, in the present invention, the occurrence frequency of the V / UV information having the lowest occurrence frequency is sequentially erased while allocating to the occurrence frequency of the adjacent V / UV information, and finally the representative V / UV of the target number is generated. UV
I took a way to reduce the number of information. Figure 3 is V / UV
When sequentially deleting the number of information (V / UV pattern) candidates,
It is a figure explaining the method of erasing the V / UV pattern with the lowest occurrence frequency. Here, the frequency of low V / UV pattern most frequency q [n] to the q [n] the frequency q of V / UV pattern adjacent each divided into q1 and q2 [nl _1] and q
Allocate to [n + l ₂ ]. The distribution of q1 and q2 depends on the size of adjacent V / UV pattern occurrences q [nl ₁ ] and q [n + l ₂ ]
It is determined by the following formula according to the proximity of the distances l ₁ and l ₂ to the V / UV pattern.

[Equation 2]

Returning to FIG. 1 again, the encoding of the spectrum envelope will be described. As the spectrum envelope, a discrete spectrum is input to the input terminal 140 as a harmonics spectrum amplitude sequence A [l], the input switching unit 141 switches the path for each frame, and one is input to the spectrum correction unit 142. The operation of the spectrum correction unit 142 will be described later. The spectrum interpolation unit 143 interpolates and generates the harmonics spectrum amplitude sequence input as the discrete spectrum amplitude sequence between the discrete spectra, and the linear prediction modeling unit 144 produces many spectrum amplitude sequences effective for modeling. In this spectrum interpolation, the harmonic spectrum amplitude is transformed into a logarithm to perform linear interpolation, and as a result, nonlinear interpolation is performed. Of course, other non-linear interpolation may be used, and an interpolation method that is well suited to the linear prediction model is desired.

The spectrum amplitude sequence whose number of samples is increased by interpolation is modeled by an autoregressive linear prediction model in the linear prediction modeling unit 144, and a gain g and a high-order (for example up to 10th) linear prediction coefficient (LPC) are obtained. Coefficient a [j], j = 1,
2,3, ..., J, J are converted to the predicted order). The gain g is converted into a logarithmic gain by the gain quantizer 145 and uniformly quantized,
The index is the spectral gain code output terminal 155.
Is output to. On the other hand, the LPC coefficient a [j] is converted into a line spectrum pair (LSP) F [j] by the modeling coefficient conversion unit 147. The LSP has a value of 0 to π, has a fixed change range, has little auditory deterioration due to linear interpolation, and is widely and generally used as a coefficient for modeling a spectrum envelope. The LPC coefficient converted into the LSP is vector-quantized by the LSP quantizer 148 using the LSP codebook 149, and the index of the codebook is output to the spectrum envelope code output terminal 156 as the spectrum envelope code.

Here, the function and purpose of the spectrum correction unit 142 will be described. When modeling a harmonics spectrum amplitude sequence with a linear prediction model, it is better to have many spectral points, but if there is a spectrum that does not fit the model, it will distort the modeling and increase the spectral error after modeling. To do. In general, the 0th-order harmonic spectrum amplitude (DC component) of voice is lower than other harmonic spectrum amplitudes, and a modeling error is likely to occur. Further, since the 0th-order harmonic spectrum amplitude is an unnecessary component at the time of voice decoding (the DC signal is hardly contained in the voice signal), it can be said that it may be adjusted and changed to a level that can be easily modeled. For the above reason, the spectrum correction unit 142 sets 0 to facilitate modeling with a linear prediction model.
Modify the following harmonics spectrum amplitude. Specifically, the value calculated by the following equation is replaced with the 0th-order harmonic amplitude.

[Equation 3] Here, H ′ ₀ is the corrected zero-order harmonics spectrum amplitude, and H ₁ and H ₂ .H ₃ are the first-order, second-order, and third-order harmonics spectrum amplitudes, respectively. Note that the zeroth-order harmonics amplitude may be replaced by using an equation other than this equation.

On the other hand, the other input harmonics spectrum amplitude sequence A [l] switched by the input switching unit 141 for each frame is the spectrum interpolation point comparison unit 154.
Entered in. In the spectrum interpolation point comparison unit 154, the interpolation gain candidates obtained by the interpolation calculation of the quantization gains of the frames before and after the input / output end of the gain delay unit 146 in the gain interpolation calculation unit 150, and the LSP interpolation calculation unit 152 in the LSP interpolation calculation unit 152. The harmonics spectrum amplitude A ′ _i [l] (l is obtained by the spectrum restoration unit 153 from the combination of the interpolated LSP coefficient candidates obtained by the interpolation calculation of the LSP coefficients of the frames before and after the input / output end of the delay unit 151. The harmonics number, i is the interpolation candidate number or index), and the input harmonics spectrum amplitude sequence A
The index i which gives the interpolation gain and the interpolation LSP coefficient with the smallest spectrum error compared with [l] is selected by the following formula, and is output to the spectrum interpolation code output terminal 157 as the spectrum interpolation code.

[Equation 4] Here, the argmin_i (x) function (the function in which i is described below argmin in the above expression) is a function that evaluates i as a parameter and returns i with the minimum x.

Further, there is a case where the above-mentioned interpolation gain candidates are inconvenient only with the interpolation candidates between the quantization gains of the preceding and following frames. For example, when there is a sharp minimum gain, the coded voice cannot reach a sufficiently low level, and unpleasant noise is generated. In order to avoid this, this problem is mitigated by adding a candidate having a minimum value or more or a maximum value of the quantization gain of the preceding and following frames to the interpolation gain candidate. Specifically, the maximum value (the value with the larger quantization gain of the preceding and following frames) +5 dB, +10 dB, the minimum value (the value with the smaller quantization gain of the preceding and following frames) -5 dB, -1
0 dB or the like is added to the interpolation gain candidates obtained by the interpolation calculation of the quantization gains of the preceding and following frames.

Next, the processing flow of the encoding unit shown in FIG. 1 according to the present invention will be described. FIG. 4 is a flow chart of encoding the voice fundamental frequency ωo. The processing starts from 701 in FIG. Voice fundamental frequency ω quantized by 702
Set o and frame number m. Next, in 703, the logarithmic pitch P is calculated, and in 704, the even / odd of m is determined. If it is even, 705 and 706 perform uniform quantization and differential quantization, respectively, to obtain a uniform quantization pitch P_u and a uniform quantization index. Index_u, differential quantization pitch P_d, and differential quantization index Index_d are calculated. Next, at 708, the difference quantization error (| P_d−P |) is determined. If it is smaller than a certain threshold Th, at 710, the pitch P [m] of the m-th frame (this is P [2n ]] Is P_d, and its index is Index [m] (this is an index in the sense of an even frame).
Let [2n] be Index_d. On the other hand, at 708 | P_d−P
When | is determined to be Th or more, the uniform quantization error (| P_u−P |) and the differential quantization error (| P_d−P |) are compared at 707, and when the uniform quantization error is small, 709, the pitch P [2n] of the m-th frame is P_u, and its index Index [2
n] as Index_u, and vice versa at 710 with pitch P [2n] P_
Let d and Index [2n] be Index_d. Further, the pitch P [2n] of the m-th frame is 711 and P [2n-
2] and the reference logarithm pitch of the differential quantization 706 is used as the reference logarithmic pitch when the next even frame is differentially quantized.

If m is an odd number in 704, the pitch interpolation candidate is selected in 713. The pitch interpolation candidates are a set of interpolation pitch candidates {Pinpol [2n-1] _i } (i = 0 , 1,2,3 ..., N-1, N are calculated in 712 as interpolation candidate points). The pitch interpolation candidate is selected by the logarithmic pitch P [2n-1] of m = 2n-1 frames.
The interpolation candidate point having the smallest absolute error value between and is selected, and its index number Index [2n-1] is set at 714 as the quantized pitch code of the odd frame. The pitch code in the case of an even frame is the index Index_d or Index_u of the quantization method selected in 709 or 710.
The index [2n] is similarly set in 714. The pitch codes for two frames, the even and odd frames set in this way, are output to 715. Using the method of encoding the fundamental frequency of speech according to the present invention shown in FIG. 4, for example,
By using 4 bits for the pitch code and 1 bit for the pitch interpolation code, it is possible to satisfactorily encode the voice fundamental frequency for 2 frames.

FIG. 5 is a flow chart of encoding V / UV information.
The processing is started from 801, and in 802, the audio fundamental frequency is set to ωo, the V / UV information value of the frame is set to v, and the frame number is set to m. At 803, the number K of bands for ωo is obtained. K = (int) ((π-ωo / 2) / (ωo × B)) where B is the number of harmonics contained in each band, which is determined in advance before encoding and is about 3 used. In 804, it is determined whether m is even or odd, and when m is an even number, in 805, the group V for the number of bands K is selected from the data 806 of the representative V / UV information value for each number of bands selected in advance.
Select the index i of the representative V / UV information value that has a value closest to the input V / UV value v in B [K] [i]} and select IndexV / UV [2n]
Then, in 807, a voiced / unvoiced code is set. When the number of processed frames is odd, voiced / unvoiced codes are decimated and not output. If the V / UV information encoding method according to the present invention shown in FIG. 5 is used, it is possible to favorably encode V / UV information for two frames by using 2 bits for a voiced / unvoiced code, for example.

FIG. 6 is a flow chart of encoding a harmonics spectrum amplitude sequence. The processing starts from 901, and 90
At 2, set the harmonics spectrum amplitude to A [l]. The frame number is set to m. In 903, it is determined whether m is even or odd. In the case of an even frame, in 904, the 0th-order harmonics spectrum amplitude A [0] is corrected by multiplying the 2nd-order harmonics spectrum amplitude A [2] by α. The method of determining the correction coefficient α is determined by the method described above. The harmonic spectrum amplitude sequence corrected in 905 is linearly interpolated with a logarithmic value to increase the number of spectrum sequences. In 906, the increased spectral sequence by interpolation is modeled and expressed as a value on the modeling curve in the linear prediction model. The gain g, which is a modeling coefficient of the linear prediction model, is uniformly quantized as a logarithmic gain at 908 to obtain a quantization gain g ′ and its index. The quantization step and the quantization range are predetermined at 907. LP, the other modeling factor
The C coefficient a [j] is converted into a line spectrum pair (LSP) and then F
[j], vector quantization is performed at 909 using the vector quantization table 910 to obtain a quantized vector F ′ [j] and its index.

The quantization gain g'from 908 is delayed by 912 for two frames, and a plurality of (M1) interpolation gain candidates between its input and output are subjected to logarithmic gain linear interpolation operation at 911. 91
Guided to 5. Also, the quantized LSP (F ′ from 909
[j]) is delayed for two frames at 913, and a plurality (M2) of interpolated LSP candidates between its input and output are subjected to LSP linear interpolation candidate calculation at 914, and the LSP interpolation candidate sequence is led to 915. At 915, the interpolation harmonics spectrum A ′ [l] is restored from the combination (maximum M1 × M2) between the input LSP interpolation candidates and gain interpolation candidates. In 916, a set of restored interpolated harmonics spectrum sequences and 903
The harmonic spectrum sequences of the odd frames input from are sequentially compared, and the combination number that gives the interpolated harmonic spectrum sequence with the smallest spectrum error is selected as the spectrum interpolation number. Further, the spectrum gain code and the spectrum envelope code shown in FIG. 1 are obtained from 908 and 909, respectively, and this processing ends at 917. By using the encoding method of the harmonics spectrum amplitude sequence according to the present invention shown in FIG. 6, the spectrum-gain code and the spectrum envelope code for two spectrum-corrected frames are converted into the spectrum gain code and the spectrum envelope for one frame. Since good coding can be performed with the code and the spectrum interpolation code, it is possible to improve the voice quality and reduce the number of coded bits.

In the above description, the input switching units 103 and 141 in FIG. 1 are described as switching for each frame for the sake of clarity. However, the switching is not limited to alternate switching for each frame. It is possible to easily realize the operation by changing the switching cycle as needed. This causes some changes in the flow charts of FIGS. 4, 5, and 6, but the relevant engineers and researchers can easily change the relevant parts. Further, in the above-described embodiment, in order to further reduce the number of encoded bits, the frame in which the voiced / unvoiced determination information is not transmitted is provided, but the voiced / unvoiced information may be transmitted for all the frames. Good. Further, in the above, the harmonics spectrum amplitude sequence is set to the gain and line spectrum pair (L
Although it is quantized by SP), it may be quantized by the gain and the linear prediction coefficient (LPC coefficient).

[0038]

As described above, according to the voice encoding method and apparatus of the present invention, the voice pitch (or the voice fundamental frequency) and the V / V of each spectrum band are set for each voice frame.
In the analysis-synthesis type speech coding method represented by speech coding parameters consisting of UV information and harmonics spectrum amplitude sequence, in the speech pitch, a frame for differential quantization or uniform quantization as a logarithmic pitch, and an interframe interpolation index are used. By encoding by switching the frame to be quantized, the number of encoded bits can be significantly reduced. In addition, the V / UV information is representatively obtained by the method according to the present invention for each number of bands determined by the range of the voice fundamental frequency.
By encoding with the V / UV value index number, the number of V / UV encoded bits can be reasonably reduced. Furthermore,
For V / UV information of a frame in which V / UV information is thinned out in frames and V / UV information is not transmitted, the number of V / UV encoded bits is further reduced by the method of analogy with the preceding and following frames described in the present invention. Can be done. Further, after correcting the 0th-order harmonics amplitude value of the speech harmonics spectrum amplitude sequence by the method described in the present invention, an autoregressive linear prediction model is created, the model coefficient is quantized and transmitted, and the model coefficient itself is quantized. Without encoding, calculate multiple interpolated modeling coefficients from the already quantized model coefficients by inter-frame interpolation, and quantize with the index number of the interpolation candidate that gives the best interpolated modeling coefficient. It is possible to significantly reduce the number of coded bits of the harmonics spectrum amplitude sequence while improving the voice quality. As described above, according to the present invention, it is possible to provide a method and device for significantly lowering the coding bit rate in the analysis and synthesis type speech coding method and device. Further, in the analysis-synthesis type speech encoding method and apparatus, the speech-synthesis type which improves the encoded speech quality by shortening the frame updating period of speech encoding and prevents the increase in the number of encoded bits is provided. It is possible to provide the voice encoding method and device.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a speech coding apparatus to which a speech coding method of the present invention is applied.

FIG. 2 is a diagram for explaining an interpolation pitch candidate creation process in an interpolation pitch candidate creation unit of the present invention.

FIG. 3 is a diagram for explaining a method of erasing voiced / unvoiced information having a low occurrence frequency according to the present invention.

FIG. 4 is a flow chart of voice pitch encoding in the voice encoding method of the present invention.

FIG. 5 is a flow chart of encoding voiced / unvoiced information in the voice encoding method of the present invention.

FIG. 6 is a flowchart of encoding a harmonic spectrum amplitude sequence in the speech encoding method of the present invention.

FIG. 7 is a configuration diagram of a voice encoding / transmission device.

FIG. 8 is a block diagram of a speech coding parameter extraction unit.

FIG. 9 is a block diagram of a parameter encoding unit in a conventional speech encoding method.

[Explanation of symbols]

102 logarithmic conversion section, 103 input switching section, 105 uniform quantization section, 108 pitch comparison section, 111 delay section, 1
12 subtraction unit, 113 difference quantization unit, 114 addition unit, 117 interpolation pitch candidate creation unit, 119 interpolation point comparison unit, 132 representative voiced / unvoiced information codebook, 133
Frame thinning section, 134 Voiced / unvoiced comparison section, 14
1 input switching unit, 142 spectrum correction unit, 143
Spectral interpolation unit, 144 Linear prediction modeling unit, 14
5 gain quantization unit, 146 gain delay unit, 147 modeling coefficient conversion unit, 148 LSP quantization unit, 149 L
SP codebook, 150 Gain interpolation calculation unit, 151
LSP delay unit, 152 LSP interpolation calculation unit, 153 spectrum restoration unit, 154 spectrum interpolation point comparison unit

─────────────────────────────────────────────────── --Continued from the front page (56) References JP-A-6-130999 (JP, A) JP-A-9-172413 (JP, A) JP-A 61-150000 (JP, A) JP-A 62- 38500 (JP, A) Thomas Eriksson, Hong-Goo Kang, Pitch Quantization in Low Bit-Rate Spec Coding, ICASSP'99, 1999, p. 489-492 Seiji Sasaki, et al., Low bit rate speech codec for commercial mobile communication using mixed excitation linear predictive coding, IEICE Transactions D-II, April 2001, Vol. J84-D-II, No. 4, p. 629-640 Teruo Kuro, et al., Low-bit-rate speech codec for commercial mobile communications using multi-band excitation harmonic linear predictive coding, IEICE Transactions D-II, 2003 1 Moon, Vol. J86-D-I I, No. 1, p. 32-41 (58) Fields investigated (Int.Cl. ⁷ , DB name) G01L 11/06 G01L 13/00 G01L 19/02 G01L 19/04 JISST file (JOIS)

Claims (11)

(57) [Claims] 1. A voice encoding method for obtaining and encoding a voice encoding parameter from a voice signal which is digitized and divided into frames of a predetermined time length, wherein the voice pitch as the voice encoding parameter is a difference. Select either quantization method or uniform quantization method to select a frame to be quantized by one of the quantization methods and one of a plurality of interpolated speech pitch candidates calculated using the quantized speech pitch of the preceding and following frames. And coding with a combination with a frame to be quantized by the index number, voiced / unvoiced information as the voice coding parameter,
The step of encoding by the index number of the closest one of the limited number of representative voiced / unvoiced information, and the harmonic spectrum amplitude sequence as the voice encoding parameter, the linear prediction coefficient ( Or a line spectrum pair derived from it) and a frame to be quantized by gain, and a plurality of interpolated linear prediction coefficient (or line spectrum pair) candidates calculated using linear prediction coefficients (or line spectrum pairs) of preceding and following frames. A speech code characterized by including a step of selecting one of a plurality of combinations of interpolation gain candidates calculated using gains of preceding and following frames, and encoding by a combination of frames to be quantized by the index number. Method. 2. The difference quantization method and the uniform quantization method in the quantization of the voice pitch are selected when the difference quantization error is less than or equal to a threshold value, and when the difference quantization error is more than a threshold value, the difference quantization method is selected. This is a process of selecting a quantization method with less quantization error, and the selection of the interpolated voice pitch candidate is performed by interpolation that is closest to the voice pitch of the current frame among a plurality of interpolated voice pitch candidates calculated from voice pitches of preceding and following frames. The voice encoding method according to claim 1, wherein the process is a process of selecting a voice pitch candidate. 3. The voice encoding method according to claim 1, wherein the quantization of the voice pitch is performed using a logarithmic pitch obtained by logarithmically converting the voice pitch. 4. The voice encoding method according to claim 1, wherein the difference quantization of the voice pitch is performed by setting a larger quantization step as the difference value increases. 5. The limited number of representative voiced / unvoiced information items are deleted from a large number of pre-acquired voiced / unvoiced information in ascending order, and the removed voiced / unvoiced information is generated. It is created by reducing the frequency to a desired limited number of voiced / unvoiced information by allocating and integrating the frequencies to adjacent voiced / unvoiced information according to the distance and size to the adjacent voiced / unvoiced information. The voice encoding method according to claim 1, wherein 6. The voiced / unvoiced information distance is represented by 0 or 1 for voiced / unvoiced sound in each voice spectrum band, and the value of 0 or 1 in each voice spectrum band is represented by a binary number in order of frequency of the voice spectrum band. 2. The speech coding method according to claim 1, wherein the speech coding method is applied to bits and expressed with a difference between the binary values. 7. The voice encoding method according to claim 1, wherein said voiced / unvoiced information is constituted by a combination of a frame for transmitting voiced / unvoiced information and a frame for omitting not to be transmitted. 8. The frame in which the voiced / unvoiced information is omitted is decoded by using voiced / unvoiced information of a frame having a large energy or an average amplitude value among the preceding and following frames. Item 7. The audio encoding method according to Item 7. 9. When modeling the harmonics spectrum amplitude sequence with a linear prediction model, a 0th order (DC component)
2. The speech coding method according to claim 1, wherein the harmonic prediction spectrum amplitude value is corrected, and then linear prediction modeling of the harmonics spectrum amplitude is performed. 10. The plurality of interpolation gain candidates include a gain obtained by equally or unequally dividing the gains of the preceding and following frames, a gain having a larger gain and a gain having a smaller gain in the preceding and following frames. The speech encoding method according to claim 1, further comprising: 11. A voice encoding device for acquiring and encoding a voice encoding parameter from a voice signal which is digitized and divided into frames of a predetermined time length, wherein a voice pitch as the voice encoding parameter is a difference. Select either quantization method or uniform quantization method to select a frame to be quantized by one of the quantization methods and one of a plurality of interpolated speech pitch candidates calculated using the quantized speech pitch of the preceding and following frames. And means for encoding by a combination of frames quantized by the index number, and voiced / unvoiced information as the voice encoding parameter,
A means for encoding by means of the index number of the closest distance from the limited number of representative voiced / unvoiced information, and the harmonic spectrum amplitude sequence as the voice encoding parameter, the linear prediction coefficient (or The frame to be quantized by the line spectrum pair derived from it and the gain, and a plurality of interpolated linear prediction coefficient (or line spectrum pair) candidates calculated using the linear prediction coefficient (or line spectrum pair) of the preceding and succeeding frames and the front and back Speech code including means for selecting any one of a plurality of combinations of interpolation gain candidates calculated using the gains of the frames and encoding with a combination of frames to be quantized by the index number. Device.