JP3292711B2 - Voice encoding / decoding 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 an audio encoding / decoding method and apparatus for encoding and decoding audio signals at a low bit rate.

[0002]

2. Description of the Related Art As a conventional technology of a low bit rate speech coding method, a 2.4 kbps LPC (Linear Predictive Codin) is used.
g: Linear predictive coding) method and 2.4kbps MELP (Mixed E)
An xcitaion Linear Prediction (mixed sound source linear prediction) method is known. These are both U.S. federal government standard audio coding systems, the former being FS-1015 (FS is Federal Stanford).
dard), and the latter was newly selected and standardized in 1996 as an improved sound quality version of FS-1015.

The 2.4 kbps LPC system and the 2.4 kbps M
There are the following references regarding the ELP method. [1] FREDERAL STANDARD 1015, "ANALOG TO DIGITAL CON
VERSION OF VOICE BY 2400 BIT / SECOND LINEAR PREDICT
IVE CODING ", November 28, 1984 [2] Federal Information Processing Standards Publi
cation, "Analog to digital conversion of voice by
2400 bit / second Mixed Excitation Linear Prediction
(MELP) ", May 28, 1998 Draft [3] L. Supplee, R. Cohn, J. Collura and A. McCree," M
ELP: The new federal standard at 2400 bps ", Proc.
ICASSP, pp.1591-1594, 1997 [4] A. McCree and T. Barnwell III, "A Mixed Excita
tion LPC Vocoder Model for Low Bit Rate Speech Codi
ng ", IEEE TRANSACTIONS ON SPEECH AND AUDIOPROCESS
ING, VOL.3, NO.4, pp.242-250, July 1995 [5] D. Thomson and D. Prezas, "SELECTIVE MODELING
OF THE LPC RESIDUAL DURING UNVOICED FRAMES: WHITE
NOISE OR PULSE EXCITATION ", Proc. ICASSP, pp.3087
-3090, 1986

First, the principle of the 2.4 kbps LPC system will be described with reference to FIGS. 10 and 11 (refer to reference [1] for details of the processing). FIG. 10 is a block diagram showing the configuration of the speech encoder of the LPC system. The framer (11) is a buffer that stores input audio samples (a1) sampled at 8 kHz after being band-limited at 100-3600 Hz and quantized with at least 12-bit accuracy. An audio sample (180 samples) is fetched every frame (22.5 ms) and output to the audio encoding processing unit as (b1). Hereinafter, a process performed for each audio encoded frame will be described.

[0005] The pre-emphasis device (12) performs high-frequency emphasis processing on (b1) and outputs a signal (c1) subjected to high-frequency emphasis processing.
The linear prediction analyzer (13) performs linear prediction analysis on (c1) using the Durbin-Levinson method, and outputs a 10th-order reflection coefficient (d1), which is spectral envelope information. The quantizer 1 (14) scalar-quantizes (d1) for each order, and outputs the resulting 41-bit output (e1) to the error correction coding / bit packing unit (19). Table 1 shows the bit allocation for each order reflection coefficient.
Shown in An RMS (Root Mean Square) calculator (15) calculates an RMS value (effective value) as level information of the signal (c1) subjected to the high-frequency emphasis processing, and outputs an RMS value (f1). Quantizer 2 (16) quantizes (f1) with 5 bits and the result is
(g1) is output to the error correction coding / bit packing unit (19).

A pitch detection / sound classifier (17) receives the output (b1) of the framing device 11 and inputs a pitch period (20 to 15).
Six samples (corresponding to 51 to 400 Hz) and sound classification information (voiced / unvoiced / transient identification information) are extracted and output as (h1) and (i1), respectively. Quantizer 3 (1
8) quantizes (h1) and (i1) together with 7 bits, and converts the result (j1) into an error correction coding / bit packing unit (19)
Output to Here, the quantization method (7-bit code (12
Assignment of pitch information and acoustic classification information to eight types of codewords) is such that codewords in which all 7 bits are 0 and codewords in which only 1 bit out of 7 bits is 1 are unvoiced.
A code word in which all bits are 1 and a code word in which only 1 bit out of 7 bits is 0 are allocated to the transient part. Other codewords are assigned to pitch period information for voiced use. The error correction coding / bit packing unit (19) packs each of the quantized information (e1), (g1), and (j1) into 54 bits / frame to form a voice coded information frame.
Output 54 bits (k1) for each frame. In the case of wireless communication, the audio information bit string (k1) is transmitted to the receiving side through a modulator and a wireless device.

Table 1 shows the bit allocation per frame. As can be seen from the table, in the error correction coding / bit packing unit (19), if the acoustic classification of the frame is not voiced (that is, if the frame is unvoiced or a transient part), the reflection coefficient of the 5th to 10th order is obtained. Instead of sending an error correction code (20 bits). The information to be error-protected in the case of unvoiced or transient part is the upper 4 bits of RMS information, 1-4.
This is the next reflection coefficient information. Also, one synchronization bit is added to each frame.

[0008]

[Table 1]

Next, the configuration of the LPC speech decoder will be described with reference to FIG. Bit separation / error correction decoder (21)
Is the 54-bit audio information bit string received for each frame
(a2) is separated for each parameter, and when the frame is unvoiced or a transient part, error correction decoding processing is performed on the corresponding bit. Then, a pitch / sound classification information bit (b2), a tenth-order reflection coefficient information bit (e2), and an RMS information bit (g2) are output. A pitch / sound classification information decoder (22) decodes the pitch / sound classification information bit (b2) and outputs a pitch period (c2) and sound classification information (d2). The reflection coefficient decoder (23) decodes the 10th-order reflection coefficient information bit (e2) and outputs a 10th-order reflection coefficient (f2). The RMS decoder (24) outputs the RMS information bits (g
2) is decoded and RMS information (h2) is output. The parameter interpolator (25) interpolates each of the parameters (c2), (d2), (f2), and (h2) in order to improve the quality of the reproduced sound, and the results (i2), (j2) ), (O2) and (r2) are output.

Next, the sound source signal (m2) is created as follows. When the interpolated sound classification information (j2) indicates voiced, the sound classification switch (28) generates a pulse sound source (k2) synchronized with the pitch period (i2) generated by the pulse sound source generator (26). When the sound classification information (j2) indicates an unvoiced voice part, the noise generator (27) operates to select the white noise (l2) generated. When the sound classification information (j2) indicates a transient portion, a pulsed sound source (k2) for a voiced portion in the frame,
White noise for unvoiced parts (pseudo-random sound source) (l2)
Works to select. Here, the boundary between the voiced portion and the unvoiced portion in the frame is determined by the parameter interpolator (25). The pitch period information (i2) for creating the pulse sound source (k2) used here uses that of the adjacent voiced sound frame. The output of the sound classification switch (28) becomes the sound source signal (m2).

The LPC synthesis filter (30) has a linear prediction coefficient
This is an all-pole filter using (p2) as a coefficient, adds spectral envelope information to the sound source signal (m2), and outputs a signal (n2) as a result. Here, the linear prediction coefficient (p2), which is the spectrum envelope information, is calculated by the linear prediction coefficient calculator (29).
Is calculated from the reflection coefficient (o2). Also, LPC
The synthesis filter (30) is configured as a 10th-order all-pole filter using a 10th-order linear prediction coefficient (p2) for voiced,
For unvoiced, it is configured as a fourth-order all-pole filter using a fourth-order linear prediction coefficient (p2). The gain adjuster (31) outputs the RM signal to the output (n2) of the LPC synthesis filter (39).
The gain is adjusted using the S information (r2), and (q2) is output. Finally, the de-emphasis unit (32) performs the reverse process on the (q2) with the pre-emphasis unit (12) described above to reproduce the reproduced sound (s
Output 2).

The problems of such an LPC system are described below (reference [4]). Problem A: In the LPC system, voiced / unvoiced / transient sections are switched for each frame over the entire frequency band. However, when a natural sound source signal is observed while divided into small frequency bands, there are a band having a voiced property and a band having an unvoiced property. Therefore, in a frame determined to be voiced in the LPC method, a component to be driven by noise is driven by a pulse, and thus a buzz sound (sounding sound) is generated. This becomes noticeable at higher frequencies. Problem B: In a transition section where voice changes from unvoiced to voiced, a sound source signal having an aperiodic pulse may be generated.
Non-periodic pulse sound sources cannot be represented in the transient part frame of the system. Therefore, tone noise is generated. As described above, in the LPC method, there is a problem that the reproduced sound has a hard-to-hear sound quality (mechanical sound quality) due to the generation of buzz sound and tone noise.

Next, the MELP system which solves the above-mentioned problems of the LPC system and improves sound quality will be described (reference documents [2]-[4]). First, the method of improving the sound quality in the MELP method will be described with reference to FIG. As shown in FIG. 3A, when natural speech is divided into bands on the frequency axis, the periodic pulse components shown in white are dominant (voiced portions) and the bands are shown in black. There is a band (unvoiced part) where the noise component is dominant. As described above, the main cause of the mechanical reproduction sound in the LPC vocoder is, as shown in FIG. 3B, the sound source is a periodic pulse component in voiced frames and a noise component in unvoiced frames over the entire frequency band. (In the transient part frame, the frame is temporally divided into voiced and unvoiced.) To solve this problem, in the MELP system, as shown in FIG.
A mixed sound source is applied by switching between voiced and unvoiced for every five frequency bands (subbands) within one frame. This method solves the problem A of the LPC method,
This has the effect of reducing the buzz sound in the reproduced sound. Also,
In order to solve the problem B of the LPC method, the decoder has a function of extracting and transmitting non-periodic pulse information, and generating a non-periodic pulse excitation on the decoder side. In addition, in order to improve the sound quality of the reproduced voice, a technique of using an adaptive spectrum enhancement filter, a pulse spread filter, and harmonics amplitude information is adopted. Table 2 summarizes the effects of each method used in the MELP method.

[0014]

[Table 2]

Next, the configuration of the 2.4 kbps MELP system will be described with reference to FIGS. 13 and 14 (for details of the processing, refer to reference [2]). FIG. 13 is a block diagram showing a configuration of the MELP speech encoder.
Framer (41) is band-limited at 100-3800Hz,
A buffer that stores input audio samples (a3) sampled at 8 kHz and quantized with at least 12-bit precision. The buffer takes in audio samples (180 samples) for each audio encoded frame (22.5 ms), and To the chemical processing department (b3)
Output as Hereinafter, a process performed for each audio encoded frame will be described.

The gain calculator (42) calculates the logarithm of the RMS value, which is the level information of (b3), and outputs the result (c3). This process is performed for the first half and the second half of the frame. That is, the logarithm of two RMS values per frame is output as (c3). The quantizer 1 (43) linearly quantizes (c3) with 3 bits for the first half and 5 bits for the second half, and divides the result (d3) into an error correction encoder / bit packer (70). Output to Linear predictive analyzer
(44) performs linear prediction analysis on (b3) using the Durbin-Levinson method, and obtains a 10th-order linear prediction coefficient (e
Output 3). The LSF coefficient calculator (45) converts the 10th-order linear prediction coefficient (e3) into a 10th-order LSF (Line Spectrum Frequencie).
s) Convert to coefficient (f3). The LSF coefficient is a feature parameter equivalent to the linear prediction coefficient. However, since the LSF coefficient has better quantization characteristics and interpolation characteristics, it has been adopted in most recent speech coding systems. Quantizer 2 (46) is a 10-order LSF
The coefficient (f3) is quantized into 25 bits by multi-stage vector quantization with four stages, and (g3) is output to the error correction coding / bit packing unit (70).

The pitch detector (54) is provided with the framing device (4).
After calculating the integer pitch period for the signal component of 1 kHz or less of the output (b3) of 1), this integer pitch period and (b3) are converted to LP
The fractional pitch period is obtained using the signal (q3) band-limited to 500 Hz or less by F (low-pass filter) (55), and (r
Output as 3). The pitch period is given as a delay amount at which the normalized autocorrelation function is maximized, and the maximum value (o3) of the normalized autocorrelation function at this time is also output. The magnitude of the maximum value of the normalized autocorrelation function is information indicating the strength of the periodicity of the input signal (b3), and is used in the aperiodic flag generator (56) (described later). The maximum value (o3) of the normalized autocorrelation function is corrected by a correlation function corrector (53) (described later), and then corrected by an error correction coding / bit packing unit (7).
It is used for voiced / unvoiced determination of all bands in 0). Here, if the corrected maximum value (n3) of the normalized autocorrelation function is equal to or smaller than the threshold value (= 0.6), it is determined that the voice is unvoiced; The quantizer 3 (57) receives the fractional pitch period (r3) from the pitch detector (54), performs logarithmic conversion, linearly quantizes it at 99 levels, and encodes the result (s3) as an error correction code. Output to the / bit packing device (70).

Four BPFs (bandpass filters) (5
8), (59), (60) and (61) output the framer (41) output (b3) of 500 to 1000 Hz, 1000 to 2000 Hz, and 2000, respectively.
~ 3000Hz, 3000 ~ 4000Hz band limited, (t3), (u3), (v
3) and (w3) are output. Four autocorrelation calculators (62),
(63), (64) and (65) are (t3), (u3) and (v3), respectively.
For (w3) and (w3), calculate the normalized autocorrelation function for the delay amount corresponding to the fractional pitch period (r3), and calculate (x3), (y
3) Output as (z3) and (a4). Next, the four voiced / unvoiced flag generators (66), (67), (68) and (69) provide thresholds (x3), (y3), (z3) and (a4), respectively, for (x3), (y3), (z3) and (a4). =
0.6) or less is judged as unvoiced, otherwise it is judged as voiced, and the flag (1 bit) indicating voiced / unvoiced is set to (b4), (c
4) Output to the correlation function corrector (53) as (d4) and (e4). Voiced / unvoiced flags (b4), (c4), (d
4) and (e4) are used to generate a mixed sound source at the decoder. The aperiodic flag generator (56) inputs the maximum value (o3) of the normalized autocorrelation function, and sets the aperiodic flag to ON if it is smaller than the threshold value (= 0.5);
The flag is set to FF, and the aperiodic flag (1 bit) (p3) is output to the error correction coding / bit packing unit (70). The non-periodic flag (p3) is used in the decoder to generate a non-periodic pulse for expressing a transient portion and a plosive sound source.

The LPC analysis filter 1 (51) is an all-zero filter using a 10th-order linear prediction coefficient (e3) as a coefficient.
The spectrum envelope information is removed from the input voice (b3), and the resulting residual signal (l3) is output. Peakiness calculator
(56) inputs the residual signal (l3) and calculates the peakiness value
Output as (m3). The peakiness value is a parameter indicating the possibility that a pulse-like component (spike) having a peak exists in a signal, and is given by the following equation from the above reference [5].

(Equation 1) Here, N the number of samples in one frame, e _n is the residual signal. Since the numerator of the above equation (1) is more susceptible to a large value than the denominator, p takes a large value when a large spike exists in the residual signal. Therefore, the higher the peakiness value, the greater the likelihood that the frame is a voiced or plosive frame with jitter that is common in transients (in these frames, the spikes (sharp peaks) will be partially reduced). But the other parts are
Because the signal is similar to white noise).

The correlation function corrector (53) calculates a peakiness value (m3)
, The maximum value (o3) of the normalized autocorrelation function and the values of the voiced / unvoiced flags (b4) and (c4) are corrected. If the peakiness value (m3) is greater than 1.34, the maximum value (o3) of the normalized autocorrelation function is set to 1.0 (indicating voiced). If the peakiness value (m3) is larger than 1.6, the maximum value (o3) of the normalized autocorrelation function is set to 1.0 (indicating voiced), and the voiced / unvoiced flags (b4) and (c4) are set to voiced. . The correlation function corrector (53) has (d4) and (e4)
Is also input, but is output without performing the correction processing. The maximum value of the corrected normalized autocorrelation function is output as (n3), and the corrected voiced / unvoiced flags (b4), (c4) and (d4), and (e4) are voiced information (f4) for each band. ) Is output.

As described above, a voiced frame or a plosive frame having jitter partially has a spike (sharp peak), but the other portions are signals having characteristics close to white noise. There is a high possibility that the autocorrelation function becomes a value smaller than 0.5 (at this time, the aperiodic flag is set to ON). Therefore, if a voiced frame or a plosive frame having jitter is detected based on the peakiness value and the normalized autocorrelation function is corrected to 1.0, voiced / unvoiced signals in the entire band in the subsequent error correction coding / bit packing unit (70) can be obtained. The voice quality is judged to be improved by using a non-periodic pulse as a sound source at the time of decoding.

Next, detection of harmonics information will be described. The linear prediction coefficient calculator (47) converts the quantized LSF coefficient (g3) output from the quantizer 2 (46) to a linear prediction coefficient, and outputs the quantized linear prediction coefficient (h3). I do. LP
The C analysis filter 2 (48) uses the input signal (b
Remove the spectral envelope component from 3) and output the residual signal (i3). The harmonic detector (49) extracts the amplitude of the tenth harmonic (harmonic component of the basic pitch frequency) in (i3), and outputs the result (j3). The quantizer 4 (50) vector-quantizes (j3) by 8 bits and outputs the index (k3) to the error correction coding / bit packing unit (70). The harmonics amplitude information corresponds to the spectrum envelope information remaining in the residual signal (i3). Thus, by sending the harmonics amplitude information, the spectral representation of the input signal can be more accurately represented during decoding, improving the nasal quality, speaker identification performance, and vowel quality in the presence of wideband noise. (Table 2).

Error correction coding / bit packing device (70)
Is set as an unvoiced frame if the maximum value (n3) of the normalized autocorrelation function after correction is equal to or smaller than the threshold value (= 0.6) as described above, and as a voiced frame otherwise, An audio information bit string is formed, and 54 bits are output as (g4) for each frame. In the case of wireless communication, the audio information bit sequence (g4) is transmitted to the receiving side through a modulator and a wireless device. In Table 3, the pitch and the entire voiced / unvoiced information are quantized by 7 bits. The method is as follows. 7-bit code (128 kinds of code words)
A code word in which all bits are 0 (or 1) and a code word in which only 1 bit out of 7 bits is 1 are unvoiced, and a code word in which 2 bits out of 7 bits are 1 is erased (erased).
Assign to Other codewords are assigned to pitch period information (output (s3) of quantizer 3 (57)) for voiced use. Voicedness for each band is (b4), (c4), (d4) and
When (e4) indicates voiced, 1 is assigned, and when unvoiced, 0 is assigned, and transmitted with 4 bits. Also, as can be seen from the table, if the frame is unvoiced, the harmonic amplitude (k
3), Instead of sending voicedness (f4) and aperiodic flag (p3) for each band, perform error correction on bits that are important for hearing,
The error correction code (13 bits) is sent. Also, one synchronization bit is added to each frame.

[0024]

[Table 3]

Next, the configuration of the MELP speech decoder will be described with reference to FIG. Bit separation / error correction decoder (8
1) Extracts the pitch and the entire voiced / unvoiced information from the 54-bit voice information bit sequence (a5) received for each frame, and if it indicates a voiceless frame, performs error protection on the corresponding bit. Perform decryption processing. Also pitch,
If the entire voiced / unvoiced information indicates an erase, each parameter is replaced with that of the previous frame. And
Pitch, as information bits of each separated parameter
Overall voiced / unvoiced information (b5), aperiodic flag (d5), harmonics amplitude index (e5), voicedness per band (g5),
LSF parameter index (j5) and gain information
(m5) is output. Here, voicedness (g5) for each band is calculated for each subband (0 to 500 Hz, 500 to 1000 Hz, 1000 to 2000 Hz, 200
0-3000Hz, 3000-4000Hz) is a 5-bit flag indicating voicedness. For voicedness of 0-500Hz, pitch,
The whole voiced / unvoiced information extracted from the whole voiced / unvoiced information is used.

A pitch decoder (82) decodes the pitch and the pitch period when the entire voiced / unvoiced information indicates voiced,
If it indicates unvoiced, 50.0 is set as the pitch period, and the decoded pitch period (c5) is output. Jitter setting device (10
2) Input the aperiodic flag (d5) and the aperiodic flag is ON
Indicates that the jitter value is 0.25, and if it indicates OFF, the jitter value is set to 0, and (g6) is output. Here, when the voiced / unvoiced information indicates unvoiced, the jitter value (g6) is 0.
Set to 25. The harmonics decoder (83) decodes and outputs the tenth-order harmonics amplitude (f5) from the index (e5) of the harmonics amplitude. The pulse sound source filter coefficient calculator (84) inputs the voicedness (g5) for each band, and sets the gain of the subband indicating voiced to 1.0 and the gain of the subband indicating unvoiced to 0. Calculate and output the coefficient (h5) of the appropriate FIR filter. The noise source filter coefficient calculator (85) inputs the voicedness (g5) for each band, and sets the gain of the subband indicating voiced to 0 and the gain of the subband indicating unvoiced to 1.0. FI like
Calculate and output the coefficient (i5) of the R filter. The LSF decoder (87) decodes and outputs a 10-order LSF coefficient (k5) from the LSF parameter index (j5). The inclination correction coefficient calculator (86) calculates the inclination correction coefficient (l5) from the 10th-order LSF coefficient (k5).
Is calculated. The gain decoder (88) decodes the gain information (m5) and outputs gain information (n5).

The parameter interpolator (89) calculates each parameter (c
5), (g6), (f5), (h5), (i5), (l5), (k5) and (n5) are linearly interpolated in synchronization with the pitch period, respectively, and (o)
5) Output (p5), (r5), (s5), (t5), (u5), (v5) and (w5). Here, the linear interpolation processing is performed by the following equation. Interpolated parameter = current frame parameter x int +
The parameter of the previous frame × (1.0−int) where the parameters of the current frame are (c5), (g6), (f5),
(h5), (i5), (l5), (k5) and (n5) respectively, and the parameters after interpolation are (o5), (p5), (r5), (s5), (t
5), (u5), (v5) and (w5) respectively. The parameters of the previous frame are (c5), (g
6), (f5), (h5), (i5), (l5), (k5) and (n5) are retained. int is the interpolation coefficient,
It is calculated by the following equation. int = to / 180 where 180 is the number of samples per audio decoded frame length (22.5 ms), to is the starting point of one pitch cycle in the decoded frame, and every time the reproduced audio of one pitch cycle is decoded It is updated by adding the pitch period. If to exceeds 180, the decoding process for that frame has ended, and 180 is subtracted from to.

The pitch period calculator (90) receives the interpolated pitch period (o5) and the jitter value (p5), and inputs the pitch period (q
5) is calculated by the following equation. Pitch period (q5) = pitch period (o5) x (1.0-jitter value (p
5) × random value) Here, the random value takes a value in the range of -1.0 to 1.0. According to the above formula, jitter is added to the unvoiced or aperiodic frame because the jitter value is set to 0.25, and no jitter is added to the periodic frame because the jitter value is set to 0. However, since the jitter value is interpolated for each pitch, there is an intermediate pitch section because it takes a range of 0 to 0.25. Generating an aperiodic pitch (pitch to which jitter has been added) based on the aperiodic flag as described above requires an irregular (aperiodic) glottis caused by a transient portion and a plosive as shown in Table 2. Expressing a pulse has the effect of reducing tone noise.

After the pitch period (q5) is converted to an integer value,
It is input to a one-pitch waveform decoder (101). The one-pitch waveform decoder (101) decodes and outputs the reproduced sound (f6) for each pitch period (q5). Therefore, all the blocks included in this block receive the pitch period (q5) and operate in synchronization with it. The pulse sound source generator (91) receives the interpolated harmonics amplitude (r5) and generates a pulse sound source (x5) having a single pulse to which the harmonics information is added. This pulse source (x5) generates one pulse in the pitch period (q5). The pulse filter (92) is an FIR filter using the interpolated pulse filter coefficient (s5) as a coefficient,
Filter the pulse sound source (x5) so that only voiced subbands are valid, and output (y5). The noise generator (94) generates white noise (a6). Noise filter (93)
Is the coefficient of the interpolated noise filter coefficient (t5).
The IR filter filters the noise source (a6) so that only unvoiced subbands are valid, and outputs (z5). The mixed sound source generator (95) adds (y5) and (z5) to generate a mixed sound source (b6). This mixed sound source has an effect of reducing buzz sound by switching between voiced / unvoiced sound source for each frequency band as shown in Table 2.

The linear prediction coefficient calculator (98) calculates a linear prediction coefficient (h6) from the interpolated 10th-order LSF coefficient (v5). The adaptive spectral enhancement filter (96) is an adaptive pole / zero filter having a coefficient obtained by performing a bandwidth extension process on the linear prediction coefficient (h6), and as shown in Table 2, sharpens the resonance of the formant. By improving the degree of approximation of the natural sound to the formant, the naturalness of the reproduced sound is improved. Further, the slope of the spectrum is corrected by using the interpolated slope correction coefficient (u5) to reduce the muffled sound, and a sound source signal (c6) as a result is output. The LPC synthesis filter (97) is an all-pole filter that uses the linear prediction coefficient (h6) as a coefficient, adds spectral envelope information to the sound source signal (c6), and outputs the resulting signal (d6). I do. The gain adjuster (99) performs gain adjustment on (d6) using the gain information (w5), and outputs (e6). Pulse spread filter
(100) is a filter for improving the approximation of the pulse sound source waveform with respect to the glottal pulse waveform of natural speech, and (e6)
Sound with improved naturalness by filtering (f6)
Is output. The effect of this pulse diffusion filter is as shown in Table 2. As described above, in the MELP method, L
Compared with the PC system, it is possible to provide a natural and intelligible reproduced sound at the same bit rate (2.4 kbps).

[0031]

Due to the explosive spread of mobile communication, it is necessary to increase the number of users accommodated, and further effective use of frequency resources is an issue. Further lowering the bit rate of the audio coding scheme is one of the essential technical issues for solving this problem. Accordingly, it is an object of the present invention to provide a speech encoding / decoding method and apparatus capable of solving the above problem A of the LPC scheme with a smaller number of bits than the 2.4 kbps MELP scheme. That is, it is an object of the present invention to provide a speech encoding / decoding method and apparatus capable of obtaining the same effect without transmitting all voiced information for each band as in the MELP method.

[0032]

In order to achieve the above object, a speech decoding method according to the present invention provides a speech information which is an output of a speech signal encoded by a speech encoder of a linear prediction analysis / synthesis system. An audio decoding method for reproducing an audio signal from a bit sequence, comprising separating and decoding spectral envelope information, voiced / unvoiced identification information, pitch period information, and gain information included in the audio information bit sequence, and extracting a spectral envelope from the spectral envelope information. The amplitude is obtained, and a band in which the value of the spectrum envelope amplitude is maximized among the bands divided on the frequency axis is determined, and the determined band and the voiced /
Based on the unvoiced identification information, a mixing ratio for mixing a pitch pulse generated based on the pitch period information with white noise is determined for each band, and a mixing ratio for each band is determined based on the determined mixing ratio. After the signal is created, a mixed sound source signal is created by adding the mixed signals of all the bands, and the reproduced sound is generated by adding the spectrum envelope information and the gain information to the mixed sound source signal. Things. As a result, the above-described problem A of the LPC scheme can be solved without transmitting additional information bits.

According to another speech decoding method of the present invention, a speech information bit string in which a speech signal is encoded by a speech encoder of a linear prediction analysis / synthesis method, which includes spectrum envelope information and voiced speech in a low frequency band. An audio decoding method for reproducing an audio signal from an audio information bit sequence in which / unvoiced identification information, voiced / unvoiced identification information of a high frequency band, pitch period information and gain information are encoded, wherein a spectrum included in the audio information bit sequence is included. Envelope information, low frequency band voiced / unvoiced identification information, high frequency band voiced / unvoiced identification information,
The pitch period information and the gain information are separated and decoded, and based on the voiced / unvoiced identification information in the low frequency band, a mixing ratio when mixing a pitch pulse generated based on the pitch period information with white noise is determined. To create a mixed signal in the low frequency band, obtain the spectrum envelope amplitude from the spectrum envelope information, and determine the band in which the value of the spectrum envelope amplitude is maximum among the bands obtained by dividing the high frequency band on the frequency axis. Based on the determined band and the voiced / unvoiced identification information of the high frequency band, for each band obtained by dividing the high frequency band, determine a mixing ratio when mixing the pitch pulse and white noise. After creating the mixed signal, a mixed signal of the high frequency band is created by adding the mixed signals of all the bands of the high frequency band, and the mixed signal of the low frequency band and the high frequency band are mixed. The mixed signal by adding to create a mixed sound source signal, is characterized in that by adding the spectral envelope information and the gain information to the mixed sound source signal to generate a reproduced speech. Thereby, the above-described LP
It is possible to solve the problem A of the C system and to further improve the reproduced voice quality.

Still another speech decoding method according to the present invention is a speech information bit sequence in which a speech signal is encoded by a speech encoder of a linear prediction analysis / synthesis method, and includes spectral envelope information and a low frequency band. A voice decoding method for reproducing a voice signal from a voice information bit sequence in which voiced / unvoiced identification information, voiced / unvoiced identification information in a high frequency band, pitch period information, and gain information are encoded, wherein the method is included in the voice information bit sequence. The spectrum envelope information, voiced / unvoiced identification information of the low frequency band, voiced / unvoiced identification information of the high frequency band, pitch period information and gain information are separated and decoded, and based on the voiced / unvoiced identification information of the low frequency band, The mixing ratio when mixing the white noise with the pitch pulse generated based on the pitch period information linearly interpolated in synchronization with the pitch period. The spectrum envelope amplitude is determined from the spectrum envelope information, and among the bands obtained by dividing the high frequency band on the frequency axis, a band in which the value of the spectrum envelope amplitude is maximized is determined. Based on voiced / unvoiced identification information of a high frequency band, a mixing ratio when mixing the pitch pulse and white noise is determined for each band obtained by dividing the high frequency band, and the spectrum envelope information and the pitch period information are determined. The gain information, the mixing ratio of the low frequency band and the mixing ratio of each band obtained by dividing the high frequency band are linearly interpolated in synchronization with a pitch cycle, and using the mixing ratio of the low frequency band after interpolation, The pitch pulse and white noise are mixed to create a mixed signal of a low frequency band, and the pitch pulse and the white noise are mixed using a mixing ratio of each band obtained by dividing the high frequency band. After creating a mixed signal for each band obtained by dividing the high frequency band by mixing white noise, a mixed signal of the high frequency band is created by adding the mixed signals of all the bands of the high frequency band, A mixed sound source signal is created by adding the mixed signal of the low frequency band and the mixed signal of the high frequency band, and the mixed sound source signal is reproduced by adding the spectrum envelope information after interpolation and the gain information after interpolation. It is characterized by generating voice. As a result, the above-mentioned problem A of the LPC method A
, And the quality of reproduced sound can be further improved.

Further, the high frequency band is divided into three bands, and the mixing ratio of each band obtained by dividing the high frequency band when the voiced / unvoiced identification information of the high frequency band indicates voiced. Is that when the value of the spectral envelope amplitude of the lowest frequency band or the second lowest frequency band is the maximum, the ratio of the pitch pulse monotonically decreases as the frequency of the divided band increases, and the highest frequency band When the value of the spectrum envelope amplitude is the largest, the ratio of the pitch pulse in the band with the second lowest frequency is made smaller than the ratio of the pitch pulse in the band with the lowest frequency, and the ratio of the pitch pulse in the band with the highest frequency is the second. The frequency is increased by a predetermined amount from the lower frequency band. Furthermore, the high frequency band is divided into three bands, and when the voiced / unvoiced identification information of the high frequency band indicates voiced, the mixing ratio of each band obtained by dividing the high frequency band is 3 The voiced intensity of the band where the value of the spectral envelope amplitude is maximum among the two bands is set to be higher than the voiced intensity of the same band when the value of the spectral envelope amplitude of the other band is maximum. It is. Furthermore, the high frequency band is divided into three bands, and when the voiced / unvoiced identification information of the high frequency band indicates unvoiced, the mixing ratio of each band obtained by dividing the high frequency band is 3 The voiced intensity of the band where the value of the spectral envelope amplitude is maximum among the two bands is set to be smaller than the voiced intensity of the same band when the value of the spectral envelope amplitude of the other band is maximum. It is.

Further, the speech coding apparatus of the present invention
A framer for inputting audio samples sampled and quantized at a predetermined sample frequency and outputting a predetermined number of audio samples for each audio encoded frame having a predetermined time length; The logarithm of the RMS value, which is the level information of the audio sample, is calculated, and the resulting logarithm RM
A gain calculator that outputs an S value, a first quantizer that linearly quantizes the logarithmic RMS value and outputs a resulting logarithmic RMS value after quantization, and a A linear prediction analyzer that performs a linear prediction analysis and outputs a linear prediction coefficient of a predetermined order that is spectrum envelope information, and outputs the linear prediction coefficient to an LSF (Line Spectrum Frequenc
ies) an LSF coefficient calculator that converts the LSF coefficients into coefficients, outputs a LSF parameter index, and a second quantizer that quantizes the LSF coefficients. A low-pass filter that filters at a cutoff frequency and outputs a low-frequency component of the input signal; and extracts a pitch period from the low-frequency component of the input signal based on a normalized autocorrelation function calculation, and calculates a pitch period and a normalized autocorrelation function. A pitch detector that outputs a maximum value, a third quantizer that performs a logarithmic conversion of the pitch period, linearly quantizes the pitch period at a first predetermined number of levels, and outputs a pitch period index obtained as a result, Enter the maximum value of the normalized autocorrelation function, and set the aperiodic flag to ON if it is smaller than a predetermined threshold;
An aperiodic flag generator that outputs the aperiodic flag when set to FF; and removes the spectral envelope information from the one-frame audio sample using the linear prediction coefficient as a coefficient. An LPC analysis filter that outputs the residual signal, a peakiness calculator that calculates and outputs a peakiness value by inputting the residual signal, and corrects the value of the maximum value of the normalized autocorrelation function by the value of the peakiness calculator. A correlation function corrector that outputs a corrected maximum value of the normalized autocorrelation function, and if the corrected maximum value of the normalized autocorrelation function is equal to or less than a predetermined threshold, it is unvoiced; A first voiced / unvoiced determiner for determining and outputting a voiced / unvoiced flag as a result thereof, and a second predetermined period for the pitch period of the frame in which the aperiodic flag indicates an aperiod. Aperiodic pitch index generator that non-uniformly quantizes by the number of bells and outputs an aperiodic pitch index, and inputs the voiced / unvoiced flag, the aperiodic flag, the pitch period index and the aperiodic pitch index, A periodic / aperiodic pitch / voiced / unvoiced information code generator for outputting a periodic / aperiodic pitch / voiced / unvoiced information code in which these are encoded with a predetermined number of bits; A high-pass filter that filters at an off frequency and outputs a high-frequency component of the input signal; and a correlation function calculator that calculates and outputs a normalized autocorrelation function at a delay time given by the pitch period from the high-frequency component of the input signal. And if the maximum value of the normalized autocorrelation function is less than or equal to a predetermined threshold, A second voiced / unvoiced discriminator that determines a voice and outputs the resulting high-frequency voiced / unvoiced flag; the logarithmic RMS value after the quantization, the LSF parameter index, the periodic / aperiodic pitch / voiced / Unvoiced information code and the high-frequency voiced / unvoiced flag, and a bit packing unit that performs bit packing for each frame and outputs a voice information bit sequence.

Furthermore, an audio decoding device according to the present invention is an audio decoding device for decoding an audio information bit sequence for each frame encoded by the above-described audio encoding device, wherein the audio information bit sequence is provided for each parameter. And the cycle /
A bit separator for outputting an aperiodic pitch voiced / unvoiced information code, a logarithmic RMS value after quantization, an LSF parameter index, and a high-frequency voiced / unvoiced flag;
An aperiodic pitch / voiced / unvoiced information code is input, and if the state of the current frame is unvoiced, the pitch period is set to a predetermined value, and the voiced / unvoiced flag is set to 0 and output. In the case of, a voiced / unvoiced information / pitch period decoder for decoding and outputting a pitch period based on a coding rule and outputting a voiced / unvoiced flag set to 1; If voiced / unvoiced information is input and the current frame indicates unvoiced or aperiodic, the jitter value is set to a predetermined value and output. If the current frame is periodic, the jitter value is set to 0 and output. A LSF decoder that decodes and outputs the LSF coefficient of the predetermined order from the LSF parameter index, a slope correction coefficient calculator that calculates and outputs a slope correction coefficient from the LSF coefficient, Decoding the log RMS value after reduction,
A gain decoder that outputs gain information, a first linear prediction coefficient calculator that converts and outputs the LSF coefficient to a linear prediction coefficient, and a spectrum envelope amplitude calculator that calculates and outputs a spectrum envelope amplitude from the linear prediction coefficient And said voiced /
The unvoiced flag, the high-frequency voiced / unvoiced flag, and the spectrum envelope amplitude are input, and the mixing ratio information of the pulse sound source and the noise sound source is determined for each of the divided bands (hereinafter referred to as “subbands”) on the frequency axis. A pulse sound source / noise source mixing ratio calculator to output, and linearly interpolate the pitch period, the mixing ratio information, the jitter value, the LSF coefficient, the tilt correction coefficient, and the gain information in synchronization with the pitch period, A parameter interpolator that outputs a pitch period after interpolation, mixing ratio information after interpolation, jitter value after interpolation, LSF coefficient after interpolation, inclination correction coefficient after interpolation, and gain information after interpolation, and a pitch after interpolation. After inputting the period and the jitter value after interpolation, add the jitter to the pitch period after interpolation, and then output the pitch period converted to an integer value (hereinafter referred to as “integer pitch period”) And a one-pitch waveform decoder that decodes and outputs the reproduced sound of the integer pitch period in synchronization with the integer pitch period. A single pulse generator that outputs a single pulse signal within, a noise generator that outputs white noise having the length of the integer pitch period, and, based on the interpolated mixing ratio information, A mixed sound source generator that mixes the single pulse signal and the white noise and then combines them to output a mixed sound source signal; and a second calculating unit that calculates a linear prediction coefficient from the interpolated LSF coefficient.
Of a linear prediction coefficient calculator, an adaptive pole / zero filter having a coefficient obtained by performing a bandwidth expansion process on the linear prediction coefficient, and a spectrum inclination correction filter having a coefficient of the inclination correction coefficient after the interpolation. An adaptive spectrum enhancement filter that filters the mixed sound source signal to output a sound source signal with an improved spectrum, and an all-pole filter that uses the linear prediction coefficient as a coefficient. An LPC synthesis filter that adds spectrum envelope information and outputs a signal to which spectrum envelope information has been added, and performs gain adjustment on the signal to which the spectrum envelope information has been added using the gain information to reproduce the signal. A gain adjuster for outputting an audio signal; and a pulse spreading process for the reproduced audio signal. It has been those with a pulse diffusion filter for outputting a reproduced audio signal.

[0038]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of a speech encoding / decoding method and apparatus according to the present invention will be described in detail with reference to FIGS. In the following, description will be made using specific numerical values, but it should be noted that the present invention can be implemented using numerical values other than the numerical values used in the following description. FIG. 1 is a block diagram of a configuration example of a speech encoder used in the speech encoding / decoding method of the present invention. In this figure, a framer (111) is a buffer that stores input audio samples (a7) sampled at 8 kHz after being band-limited at 100-3800 Hz and quantized with at least 12-bit precision, Captures audio samples (160 samples) for each audio encoded frame (20 ms) and sends it to the audio encoding processing unit
Output as (b7). Hereinafter, a process performed for each audio encoded frame will be described.

The gain calculator (112) calculates the logarithm of the RMS value, which is the level information of the input voice (b7), and obtains the result.
(c7) is output. The first quantizer (referred to as "quantizer 1") (113) linearly quantizes (c7) with 5 bits, and outputs the result (d7) to the bit packing device (125). The linear prediction analyzer (114) performs linear prediction analysis on (b7) using the Durbin-Levinson method, and outputs a 10th-order linear prediction coefficient (e7) that is spectrum envelope information. LSF coefficient calculator (115)
Calculates the 10th-order LSF (Line Spe
ctrum Frequencies).

The second quantizer (quantizer 2) (116)
The configuration is such that the memoryless vector quantization and the prediction (memory) vector quantization using the multistage vector quantization with three stages are switched and used, whereby the 10th-order LSF coefficient (f7) is quantized by 19 bits, and the result is Is output to the bit packing unit (125). For example, the quantizer 2 (116) converts the input 10-order LSF coefficient (f7) into a 7, 6, 5 bit three-stage predictor in the same manner as the 7, 6, 5 bit three-stage memoryless vector quantizer. The data is input to the vector quantizer, quantized, and one of the outputs is selected based on the result of the distance calculation with the input (f7), and is output together with a switch bit (1 bit) indicating which is selected. Note that such a quantizer is described in T. Eriksson, J. Lind
en and J. Skoglund, "EXPLOITING INTERFRAME CORRELA
TION IN SPECTRAL QUANTIZATIONA STUDY OF DIFFERENT
MEMORY VQ SCHEMES ", Proc. ICASSP, pp.765-768, 19
95.

The LPF (low-pass filter) (120) filters the input voice (b7) at a cutoff frequency of 1000 Hz and outputs a component (k7) of 1000 Hz or less. Pitch detector (1
In (21), the pitch period is obtained from (k7) and output as (m7). The pitch period is given as a delay amount at which the normalized autocorrelation function is maximized, and the maximum value (17) of the normalized autocorrelation function at this time is also output. The magnitude of the maximum value of the normalized autocorrelation function is information indicating the strength of the periodicity of the input signal (b7), and is used by the aperiodic flag generator (122) (described later). Also, the maximum value (l7) of the normalized autocorrelation function is
After being corrected by the correlation function corrector (119) (described later), the first voiced / unvoiced determiner (voiced / unvoiced determiner 1) (1
Used for voiced / unvoiced determination in 26). Where,
If the maximum value (j7) of the corrected normalized autocorrelation function is equal to or smaller than a threshold value (for example, 0.6), it is determined that the voiced voice is unvoiced; otherwise, voiced / unvoiced flag (s7) is output. You. The voiced / unvoiced flag (s7) corresponds to voiced / unvoiced identification information in a low frequency band.

A third quantizer (quantizer 3) (123) inputs the pitch period (m7), performs logarithmic conversion, linearly quantizes the 99 levels, and converts the resulting pitch index (o7) into a period. / Aperiodic pitch and voiced / unvoiced information code generator (1
Output to 27). FIG. 3 shows the relationship between the pitch period (taken in the range of 20 to 160 samples) as an input to the quantizer 3 (123) and the index value (taken in the range of 0 to 98) as its output. The aperiodic flag generator (122) inputs the maximum value (17) of the normalized autocorrelation function, and sets the aperiodic flag to ON if it is smaller than a threshold value (for example, 0.5); otherwise, sets it to OFF. And the aperiodic flag (1 bit) (n
7) with an aperiodic pitch index generator (124) and a periodic / aperiodic pitch and voiced / unvoiced information code generator (1).
Output to 27). Here, if the aperiodic flag (n7) is ON, it means that the current frame is a sound source having aperiodicity.

The LPC analysis filter (117) is an all-zero filter using a 10th-order linear prediction coefficient (e7) as a coefficient.
The spectrum envelope information is removed from the input speech (b7), and the resulting residual signal (h7) is output. Peakiness calculator
(118) receives the residual signal (h7), calculates the peakiness value, and outputs it as (i7). The peakiness value is calculated using the same method as described in the MELP method. If the peakiness value (i7) is larger than a predetermined value (for example, 1.34), the correlation function corrector (119) sets the maximum value (l7) of the normalized autocorrelation function to 1.0 (indicating voiced), and j7) is output. When the value is equal to or less than the predetermined value, (l7) is changed to (j
Output as it is as 7).

The peakiness value calculation and correlation function correction processing described above detects a voiced frame or a plosive frame having jitter and corrects the maximum value of the normalized autocorrelation function to 1.0 (a value indicating voiced). This is the process to perform.
Voiced or plosive frames with jitter have some spikes (sharp peaks), but the rest of the signal is close to white noise, so the normalized self-corrected The correlation function is more likely to be smaller than 0.5 (that is, it is more likely that the aperiodic flag is set to ON). On the other hand, the peakiness value increases. Therefore, if a voiced frame or a plosive frame having jitter is detected based on the peakiness value and the normalized autocorrelation function is corrected to 1.0, the voiced / unvoiced determination in the voiced / unvoiced determination unit 1 (126) is determined to be voiced. Since the aperiodic pulse is used as a sound source during decoding, the voice quality of a voiced frame having a jitter or a plosive frame is improved.

Next, the aperiodic pitch index generator (1
24) and the periodic / aperiodic pitch and voiced / unvoiced information code generator (127) will be described. By transmitting periodic / non-periodic identification information using these, and switching between periodic / non-periodic pulses in a decoder to be described later, tone-like noise, which is the problem B of the LPC method, can be reduced.

Aperiodic pitch index generator (124)
Outputs a non-uniform quantization of the pitch period (m7) in the aperiodic frame at 28 levels and outputs an index (p7). The processing content of the non-uniform quantization will be described. First, the voiced / unvoiced flag (s7) is voiced, and the aperiodic flag (n7) is
FIG. 4 shows the result of examining the frequency of the pitch period for a frame having N (corresponding to a voiced frame or a plosive frame having a jitter in the transient portion), and FIG. 5 shows the cumulative frequency. . These are the results of measurements on a total of 112.12 [s] (5606 frames) of audio data composed of four men and women (6 audio samples / one each). Frames satisfying the above conditions (voiced / unvoiced flag (s7) is voiced and aperiodic flag (n7) is ON) are 42 out of 5606 frames.
There were 5 frames. From FIGS. 4 and 5, it can be seen that the pitch period distribution in frames satisfying the condition (hereinafter referred to as aperiodic frames) is concentrated at about 25 to 100. Therefore, if non-uniform quantization based on the frequency (appearance frequency) is performed, that is, if the pitch period is larger, the finer the pitch period is, and the smaller the pitch period, the coarser the quantization, the higher the transmission efficiency.

As will be described later, on the decoder side, the pitch period of the aperiodic frame is calculated by the following equation. Pitch period of aperiodic frame = transmitted pitch period x
(1.0 + 0.25 × random value) In the above expression, the transmitted pitch period is a pitch period transmitted by an index output from the aperiodic pitch index generator (124), and is expressed as (1.0 + 0. By multiplying by 25 × random value, jitter is added for each pitch period. Therefore, the larger the pitch period is, the larger the amount of jitter is, so that rough quantization is allowed.

Table 4 shows an example of the quantization table for the pitch period of the aperiodic frame based on the above idea. In the table, the range of the input pitch period from 20 to 24 is one level,
13 levels (2 step widths) in the range of ~ 50, 9 levels (5 step widths) in the range of 51 to 95, 4 levels (10 step widths) in the range of 96 to 135, and 1 level in the range of 136 to 160 Quantize and output index (non-period 0 to 27). While normal pitch cycle quantization requires 64 levels or more, this non-periodic frame pitch cycle quantization can be performed at 28 levels by considering the frequency and decoding method. Becomes

[0049]

[Table 4]

The periodic / aperiodic pitch and voiced / unvoiced information code ゛ generator (127) includes a voiced / unvoiced flag (s7), an aperiodic flag (n7), a pitch index (o7), and an aperiodic pitch index (p7). ) Is input and 7 bits (128 levels)
Output a periodic / aperiodic pitch / voiced / unvoiced information code (t7). The processing here will be described below. When the voiced / unvoiced flag (s7) indicates unvoiced, a 7-bit code (12
(7 types of codewords are allocated). If the flag indicates voiced, the remaining codewords (127 types) are pitch index (o7) or aperiodic pitch index (p) based on the aperiodic flag (n7).
Assign to 7). When the aperiodic flag (n7) is ON, the aperiodic pitch index (p7) (aperiods 0 to 27) is assigned to codewords in which 1 bit or 2 bits are 1 in 7 bits (total 28 types). Other code words (99 types) are assigned to a periodic pitch index (o7) (periods 0 to 98).

Based on the above, periodic / aperiodic pitch / voiced /
Table 5 shows a generation table of the unvoiced information code. Normally, when an error occurs in voiced / unvoiced information due to a transmission error and an unvoiced frame is erroneously decoded as a voiced frame, the quality of the reproduced voice is significantly deteriorated because a periodic sound source is used. As described above, by assigning the aperiodic pitch index (p7) (aperiods 0 to 27) to codewords in which 1 bit or 2 bits are 1 in 7 bits (total 28 types),
Even if the unvoiced code word (0x0) is erroneous by one or two bits due to a transmission error, an excitation signal is generated by an aperiodic pitch pulse, so that the effect of the transmission error can be reduced. Alternatively, all 1s (0x7F) may be assigned to unvoiced codewords, and codewords in which 1 or 2 bits out of 7 bits are 0 may be assigned to the aperiodic pitch index. In the MELP method described above, one bit is used for transmitting the aperiodic flag. However, by using this method, it is not necessary, and the number of transmission bits can be reduced.

[0052]

[Table 5]

The HPF (high pass filter) (128) filters the input voice (b7) at a cutoff frequency of 1000 Hz and outputs a high frequency component (a component of 1000 Hz or more) (u7). The correlation function calculator (129) calculates and outputs a normalized autocorrelation function (v7) with respect to (u7) in a delay amount given by the pitch period (m7). The second voiced / unvoiced determiner (voiced / unvoiced determiner 2) (130) is a normalized autocorrelation function (v7)
If is less than or equal to a threshold value (for example, 0.5), it is determined that the voice is unvoiced. This high-frequency voiced / unvoiced flag corresponds to voiced / unvoiced identification information in a high frequency band.

The bit packing unit (125) includes a quantized RMS value (gain information) (d7), an LSF parameter index (g7), a periodic / aperiodic pitch voiced / unvoiced information code (t7), and a high-frequency voiced voice. Enter the / silent flag (w7)
32-bit audio information bit string per frame (20 ms)
(q7) is output (Table 6). As a result, in the embodiment shown here, a speech coding speed of 1.6 kbps can be realized.
In the present embodiment, the harmonic amplitude information is not transmitted as in the MELP method described above. The reason is as follows. Since the speech coding frame length is shortened to 20 ms (22.5 ms in the MELP system), the period for extracting the LSF parameter is shortened, and the accuracy of spectrum expression is improved. Therefore, no harmonic amplitude information is required. In addition, here, HPF (128), correlation function calculator
(129) and a voiced / unvoiced decision unit 2 (130) are provided to transmit the high-frequency voiced / unvoiced flag (w7). As described later, the high-frequency voiced / unvoiced flag (w7) is transmitted. Need not be sent.

[0055]

[Table 6]

Next, an embodiment of a speech decoder to which the speech decoding method of the present invention is applied, which decodes a speech information bit string encoded by the above speech encoder and reproduces a speech signal, will be described with reference to FIG. This will be described with reference to FIG. In FIG. 2, a bit separator (131) separates a 32-bit audio information bit string (a8) received for each frame for each parameter, and generates a periodic / aperiodic pitch / voiced / unvoiced information code (b8),
High frequency voiced / unvoiced flag (f8), gain information (m8) and LS
Output the F parameter index (h8). The voiced / unvoiced information / pitch period decoder (132)
The voiced / unvoiced information code (b8) is input, and based on the table shown in Table 5, which of unvoiced / periodic / aperiodic is determined. If unvoiced, the pitch period (c8) is determined. (For example, 50) and the voiced / unvoiced flag (d8) is set to 0
Set to and output. In the case of periodic and non-periodic, the pitch period (c8) is decoded (in the case of non-periodic, use Table 4 above) and output, and the voiced / unvoiced flag (d8) is set to 1.
Set to 0 and output.

The jitter setting unit (133) receives the periodic / non-periodic pitch / voiced / unvoiced information code (b8) and determines which of unvoiced / periodic / non-periodic based on Table 5. If the signal indicates unvoiced or aperiodic, the jitter value (e8) is set to a predetermined value (for example, 0.25) and output. If it indicates periodicity, the jitter value (e8) is set to 0 and output. The LSF decoder (138) decodes and outputs the tenth-order LSF coefficient (i8) from the LSF parameter index (h8).
The inclination correction coefficient calculator (137) calculates an inclination correction coefficient (j8) from the tenth-order LSF coefficient (i8). The gain decoder (139) decodes the gain information (m8) and outputs the gain information (n8).
First linear prediction coefficient calculator (linear prediction coefficient calculator 1) (1
36) converts the LSF coefficient (i8) into a linear prediction coefficient and outputs a linear prediction coefficient (k8). Spectral envelope amplitude calculator (1
35) is the spectral envelope amplitude (l8) from the linear prediction coefficient (k8).
Is calculated. Here, the voiced / unvoiced flag (d8) and the high-frequency voiced / unvoiced flag (f8) are the voiced / unvoiced flag in the low frequency band, respectively.
The voiceless identification information corresponds to voiced / unvoiced identification information in a high frequency band.

Next, a pulse sound source / noise sound source mixture ratio calculator
The configuration of (134) will be described. FIG. 6 is a diagram showing a configuration of the pulse sound source / noise sound source mixture ratio calculator (134).
Voiced / unvoiced flag (d8), spectral envelope amplitude at
(l8) and the high-frequency voiced / unvoiced flag (f8) are input, and the mixing ratio (g8) of each band (sub-band) is determined and output. In this embodiment of the speech decoding method of the present invention, the signal is divided into four bands on the frequency axis, and a mixing ratio between a pulse sound source and a noise sound source is determined for each band to generate a mixed signal. Are used. Subband 1 (0 to 1000H)
z), sub-band 2 (1000-2000 Hz), sub-band 3 (2
000-3000Hz) and sub-band 4 (3000-4000Hz)
Set. Subband 1 corresponds to a low frequency band, and subbands 2, 3, and 4 correspond to high frequency bands.

Subband 1 voiced intensity setting unit (160) in FIG.
Inputs the voiced / unvoiced flag (d8) and sets the voiced strength (a10) of subband 1. Here, the voiced / unvoiced flag
If (d8) is 1.0, the voiced intensity (a10) is set to 1.0, and if the voiced / unvoiced flag (d8) is 0, the voiced intensity (a10) is set to 0.
The sub-band 2, 3, and 4 average amplitude calculator (161) receives the spectrum envelope amplitude (18), calculates the average value of the spectrum envelope amplitude in the sub-bands 2, 3, and 4, and calculates (b
10), output as (c10) and (d10). The subband selector (162) inputs (b10), (c10), and (d10), and receives the subband number (e
Output 10). The subband 2, 3, and 4 voiced strength table (for voiced) (163) is composed of three three-dimensional vectors ((f101),
(f102) and (f103)), and each three-dimensional vector is composed of the voiced intensities of the subbands 2, 3, and 4 at the time of the voiced frame. The first switch (switch 1) (165) selects and outputs one vector (h10) from the three three-dimensional vectors according to the subband number (e10).

Subbands 2, 3, 4 Voiced strength table
(For unvoiced) (164) is also a three-dimensional vector ((g10
1), (g102), and (g103)), and the respective three-dimensional vectors are subbands 2, 3, and 4 at the time of unvoiced frames.
Is composed of voiced intensities. The second switching device (switching device 2) (166) performs three 3 switching operations according to the subband number (e10).
One vector (i10) is selected and output from the dimensional vector. The third switch (switch 3) (167) is high-frequency voiced /
Enter the unvoiced flag (f8) and if it indicates voiced (h1
(0) is selected (i10) to indicate unvoiced and output as (j10). As described above, the high-frequency voiced / unvoiced flag (w
If 7) is not sent, a voiced / unvoiced flag (d8) may be used instead of the high-frequency voiced / unvoiced flag (f8).

The mixing ratio calculator (168) receives the voiced intensity (a10) of subband 1 and the voiced intensity (j10) of subbands 2, 3, and 4, and outputs the mixing ratio (g8) of each subband. . This mixing ratio (g8) indicates the ratio of the pulse sound source in each subband.
1_p, sb2_p, sb3_p, sb4_p, and sb indicating the ratio of noise sources
1_n, sb2_n, sb3_n, and sb4_n (where,
In sbx_y, x indicates a subband number, and when y is p, a pulsed sound source and when y is n, a noise source. sb1_p, sb2_
As p, sb3_p, and sb4_p, the voiced intensity (a1
0), the values of the voiced intensities (j10) of the subbands 2, 3, and 4 are used as they are. For sbx_n (x = 1,... 4), sbx_n = (1.0−sbx_p) (x = 1,..., 4) is set.

Next, a method of determining the subband 2, 3, and 4 voiced strength table (for voiced) (163) will be described.
The values in the table are determined based on the voiced strength measurement results of subbands 2, 3, and 4 in the voiced frame shown in FIG. The measurement method shown in FIG. The average value of the spectral envelope amplitude in each of the subbands 2, 3, and 4 is calculated for each frame (20 ms) with respect to the input voice, and a group of frames (represented by fg_sb2) in which the subband 2 has the maximum value, and a subband 3 Group of frames that maximizes it (fg
_sb3) and a group of frames (represented as fg_sb4) in which the subband 4 has the maximum value. Next, the audio frame belonging to the frame group fg_sb2 is divided into subband signals corresponding to subbands 2, 3, and 4, and a normalized autocorrelation function in a pitch cycle is obtained for each subband signal. The average was determined.

The horizontal axis in FIG. 8 indicates the subband number. Since the normalized autocorrelation function is a parameter indicating the strength of the periodicity of the input signal, that is, the strength of voicedness, it means voiced strength. The vertical axis in FIG. 8 indicates the voiced intensity (normalized autocorrelation) of each subband signal. The curve marked with ◆ in the figure is fg
The results measured for _sb2 are shown. Similarly, the result measured for the frame group fg_sb3 is indicated by a curve indicated by ●, and the result measured for the frame group fg_sb4 is indicated by a curve indicated by ○. The input audio signal used in this measurement is composed of audio from the audio database CD-ROM and audio recorded from FM broadcast.

FIG. 8 shows the following tendency. In the frame where the average value of the spectral envelope amplitude in the subband 2 or 3 is maximum (marked with ◆ and circle),
The voiced intensity decreases monotonically as the frequency of the subband increases. In the frame where the average value of the spectral envelope amplitude in the subband 4 is the maximum (marked by ○), the voiced intensity does not monotonously decrease as the frequency of the subband increases, and the voiced intensity of the subband 4 becomes relatively strong. Also, the voiced intensity of the subbands 2 and 3 becomes weaker (compared to the case where the average value of the spectral envelope amplitude in the subbands 2 and 3 becomes maximum (◆ and ●). The voiced intensity of the subband 2 of the frame (marked by a triangle) in which the average value of the spectral envelope amplitude of the subband 2 is the maximum is
It becomes larger than the voiced intensity of subband 2 in the mark and the mark. Similarly, the sub-band 3 of the frame (marked by ●) in which the average value of the spectral envelope amplitude of the sub-band 3 is maximum
Is larger than the voiced intensity of the sub-band 3 at the mark Δ and the mark ○. Similarly, the frame in which the average value of the spectral envelope amplitude of subband 4 is the maximum (○
The voiced intensity of the sub-band 3 of the mark (印) is larger than the voiced intensity of the sub-band 4 at the mark Δ and the mark ●.

Accordingly, the three-dimensional vector (f101) is added to the subband 2, 3, and 4 voiced strength table (for voiced) (163) in FIG.
The value of the voiced intensity of the curve marked with ◆ is stored, the value of the voiced intensity of the curve marked with と し て is stored as (f102), and the value of the voiced intensity of the curve marked with と し て is stored as (f103), and the average value of the spectral envelope amplitude is stored. Based on the sub-band number indicated by the sub-band number (e10) that maximizes, the appropriate voiced strength can be set according to the spectral envelope amplitude if selected. Table 7 shows subbands 2, 3,
4 shows the contents of the voiced strength table (for voiced) (163).

[Table 7]

Subband 2, 3, 4 Voiced Strength Table
(For unvoiced) (164) was determined based on the voiced intensity measurement results of subbands 2, 3, and 4 in the unvoiced frame shown in FIG. The method of measurement and the method of determining the contents of the table in the figure are exactly the same as those of the above-described voiced frame. FIG. 9 shows the following tendency. The voiced intensity of the subband 2 of the frame (marked by a triangle) in which the average value of the spectral envelope amplitude of the subband 2 is the maximum is
It becomes smaller than the voiced intensity of the subband 2 in the mark and the mark. Similarly, the sub-band 3 of the frame (marked by ●) in which the average value of the spectral envelope amplitude of the sub-band 3 is maximum
Is smaller than the voiced intensity of the sub-band 3 at the mark Δ and the mark ○. Similarly, the frame in which the average value of the spectral envelope amplitude of subband 4 is the maximum (○
The voiced intensity of the sub-band 3 of the mark () is smaller than the voiced intensity of the sub-band 4 at the mark Δ and the mark ●. Table 8 shows the contents of the table.

[Table 8]

Returning to FIG. 2, the parameter interpolator (140)
Are linearly synchronized with the pitch period for each parameter of the pitch period (c8), the jitter value (e8), the mixing ratio (g8), the slope correction coefficient (j8), the LSF coefficient (i8), and the gain information (n8). Interpolated, interpolated pitch period (o8), jitter value (p8),
The mixing ratio (r8), the inclination correction coefficient (s8), the LSF coefficient (t8), and the gain information (u8) are output. The linear interpolation process here is
The following equation is used. Interpolated parameter = current frame parameter x int +
Parameter of previous frame × (1.0−int) Here, parameters of current frame are (c8), (e8), (g8),
(j8), corresponding to (i8) and (n8), the parameters after interpolation correspond to (o8), (p8), (r8), (s8), (t8) and (u8) . The parameters of the previous frame are (c8), (e8), (g8), (j8), (i8) and (n
Given by holding 8). int is an interpolation coefficient, which is obtained by the following equation. int = to / 160.0 Here, 160.0 is the number of samples per audio decoding frame length (20 ms), to is the starting sample point of one pitch period in the decoding frame, and every time the reproduced audio of one pitch period is decoded. Is updated by adding the pitch period to the data. When to exceeds 160, it means that the decoding process of the frame has been completed, and 160 is subtracted from to. Here, if the interpolation coefficient int is fixed to 1.0, the linear interpolation processing synchronized with the pitch cycle is not performed.

The pitch period calculator (141) receives the interpolated pitch period (o8) and jitter value (p8),
(q8) is calculated by the following equation. Pitch period (q8) = pitch period (o8) x (1.0-jitter value (p
8) × random value) Here, the random value takes a value in the range of -1.0 to 1.0. This pitch period (q8) has a decimal number, but is rounded off and converted to an integer. In the following, the pitch period (q8) converted to an integer is
Expressed as an integer pitch period (q8). From the above equation, the jitter value for unvoiced or aperiodic frames is a predetermined value (here, 0.2
Jitter is added because it is set to 5), and no jitter is added in a complete periodic frame because the jitter value is set to zero. However, since the jitter value is interpolated for each pitch, there is a pitch section in which an intermediate amount of jitter is added because it takes a range of 0 to 0.25. The generation of the aperiodic pitch (pitch to which jitter is added) as described above represents an irregular (aperiodic) glottal pulse generated by a transient portion and a plosive as described in the description of the MELP system. This has the effect of reducing tone noise.

The one-pitch waveform decoder (150) decodes and outputs the reproduced voice (b9) for each integer pitch period (q8). Therefore,
All blocks included in this block receive an integer pitch period (q8) and operate in synchronization with it. The pulse generator (142) outputs a single pulse signal (v8) within an integer pitch period (q8) period. Noise generator (143) is an integer pitch period
A white noise (w8) having a length of (q8) is output. The mixed sound source generator (144) outputs a mixed sound source signal (x8) by mixing the single pulse signal (v8) and the white noise (w8) based on the mixing ratio (r8) of each subband after interpolation. .

FIG. 7 shows the structure of the mixed sound source generator (144). First, a process of generating the mixed signal (q11) of subband 1 will be described. LPF (Low Pass Filter) 1 (170)
Limits the band of single pulse signal (v8) from 0 to 1kHz (a11)
Is output. LPF2 (171) converts white noise (w8) to 0 to 1 kHz
And (b11) is output. The multiplier 1 (178) and the multiplier 2 (179) add mixing ratio information (r8) to (a11) and (b11), respectively.
Are multiplied by sb1_p and sb1_n, and (i11) and (j11) are output. Adder 1 (186) adds (i11) and (j11) and outputs a mixed signal (q11) of subband 1. Similarly, the mixed signal (r11) of the subband 2 has a BPF (bandpass filter) 1 (172), a BPF 2 (173), a multiplier 3 (180), a multiplier 4 (181), and an adder 2 (189). ). Similarly, the mixed signal (s11) of the sub-band 3 has the BPF3 (17
4), BPF4 (175), multiplier 5 (182), multiplier 6 (183),
And the adder 3 (190). Subband 4
Similarly, the mixed signal (t11) of (1) is obtained using an HPF (high pass filter) 1 (176), an HPF 2 (177), a multiplier 7 (184), a multiplier 8 (185), and an adder 4 (191). Made. The adder 5 (192) outputs the mixed signals (q11), (r1
1), (s11) and (t11) are added to synthesize a mixed sound source signal (x8).

In FIG. 2, a second linear prediction coefficient calculator (linear prediction coefficient calculator 2) (147) converts the LSF coefficient (t8) after interpolation into a linear prediction coefficient, and Is output. Adaptive spectral enhancement filter (1
45) is dependent on an adaptive pole / zero filter having a coefficient obtained by performing a bandwidth extension process on the linear prediction coefficient (b10) and a spectrum inclination correction filter having a coefficient of the inclination correction coefficient (s8) after the interpolation. Table 2
As shown in FIG. 5, the resonance of the formant is sharpened, and the approximation of the natural sound to the formant is improved, thereby improving the naturalness of the reproduced sound. Further, the slope of the spectrum is corrected by a spectrum tilt correction filter using the interpolated tilt correction coefficient (s8) as a coefficient to reduce muffled sound. The mixed sound source signal (x8) is filtered by the adaptive spectrum enhancement filter (145), and the result (y8) is output.

The LPC synthesis filter (146) is an all-pole filter that uses the linear prediction coefficient (b10) as a coefficient, adds spectral envelope information to the sound source signal (y8), and generates a signal (z8 ) Is output. The gain adjuster (148)
The gain is adjusted using the interpolated gain information (u8) for (z8), and (a9) is output. Pulse diffusion filter (149)
Is a filter that performs pulse diffusion processing to improve the approximation of the pulse sound source waveform to the glottal pulse waveform of the natural voice, and outputs a reproduced voice (b9) with improved naturalness by filtering (a9) . The effect of this pulse diffusion filter is as shown in Table 2.

In the above description, the sub-band in which the average value of the spectral envelope amplitude of each sub-band is maximum is determined, and the mixing ratio is determined accordingly. And not
Other values may be used. Further, the speech encoding device and speech decoding device of the present invention described above
(Digital signal processor). Further, as a control signal of the switch 3 (167) in the pulse sound source / noise sound source mixture ratio calculator,
Voiced / unvoiced flag instead of high-frequency voiced / unvoiced flag (f8)
When (d8) is used, the conventional system (LPC system) can be used as it is as the audio encoder. Furthermore, the number of quantization levels, the number of bits of codewords, the length of a speech coding frame, the order of linear prediction coefficients, the LSF coefficients, and the cutoff frequency of each filter are used in the above-described embodiments. The value is not limited to the above value, and a value according to each case can be adopted.

[0074]

As described above, by using the speech encoding / decoding method and apparatus of the present invention, the quality of the conventional system (LPC system) is deteriorated.
The zz sound can be reduced, the sound quality of the reproduced sound can be improved, and the encoding speed can be lower than that of the conventional method (MELP method). Therefore, when used for wireless communication,
It is possible to improve the frequency use efficiency.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an embodiment of a speech encoder to which a speech encoding method according to the present invention is applied.

FIG. 2 is a block diagram showing a configuration of an embodiment of a speech decoder to which the speech decoding method of the present invention is applied.

FIG. 3 is a diagram for explaining a relationship between a pitch period and an index.

FIG. 4 is a diagram for explaining a frequency of a pitch cycle.

FIG. 5 is a diagram for explaining a cumulative frequency of a pitch cycle.

FIG. 6 is a block diagram showing a configuration example of a pulse sound source / noise source mixture ratio calculator in one embodiment of the speech decoder of the present invention.

FIG. 7 is a block diagram illustrating a configuration example of a mixed sound source generator according to an embodiment of the speech decoder of the present invention.

FIG. 8 Voiced intensity of subbands 2, 3, and 4 (when voiced)
FIG.

[FIG. 9] Voiced intensity of subbands 2, 3, and 4 (when unvoiced)
FIG.

FIG. 10 is a diagram illustrating a configuration of a conventional (LPC) speech encoder.

FIG. 11 is a diagram illustrating a configuration of a conventional (LPC) speech decoder.

FIG. 12 is a diagram illustrating spectra of the LPC system and the MELP system.

FIG. 13 is a diagram illustrating a configuration of a conventional (MELP) speech encoder.

FIG. 14 is a diagram showing a configuration of a conventional system (MELP) speech decoder.

[Explanation of symbols]

111 framer, 112 gain calculator, 113
Quantizer, 114 linear prediction analyzer, 115 LSF coefficient calculator, 116 quantizer, 117 LPC analysis filter, 118 peakiness calculator, 119 correlation function corrector, 120 low-pass filter, 121 pitch detector, 122 aperiodic flag Generator, 123 quantizer,
124 aperiodic pitch index generator, 125 bit packing device, 126 voiced / unvoiced decision device, 127
Periodic / aperiodic pitch and voiced / unvoiced information code generator, 128 high-pass filter, 129 correlation function calculator, 130 voiced / unvoiced determiner, 131 bit separator, 132 voiced / unvoiced information / pitch periodic decoder, 13
3 Jitter setting unit, 134 pulse sound source / noise source mixing ratio calculator, 135 spectrum envelope amplitude calculator, 136
Linear prediction coefficient calculator, 137 tilt correction coefficient calculator,
138 LSF decoder, 139 gain decoder, 140
Parameter interpolator, 141 pitch period calculator, 14
2 pulse source generator, 143 noise generator, 144
Mixed sound source generator, 145 Adaptive spectrum enhancement filter, 146 LPC synthesis filter, 147
Linear prediction coefficient calculator, 148 gain adjuster, 149
Pulse spread filter, 160 sub-band 1 voiced intensity setter, 161 sub-band 2, 3, 4 average amplitude calculator, 1
62 subband selector, 163 subbands 2, 3,
4 voiced strength table (for voiced), 164 subbands 2, 3, 4 voiced strength table (for unvoiced), 165 to 16
7 Switcher, 168 Mixing ratio calculator, 170, 171
LPF, 172-175 BPF, 176,177
HPF, 178-185 Multiplier, 186, 189-1
92 Adder

Continuation of the front page (56) References JP-A-2000-267700 (JP, A) JP-A-11-143499 (JP, A) JP-A-10-20892 (JP, A) JP-A-7-295593 (JP, A A) JP-A-4-116700 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/06

Claims (8)

(57) [Claims] 1. An audio decoding method for reproducing an audio signal from an audio information bit string which is an output obtained by encoding an audio signal by a linear predictive analysis / synthesis audio encoder, wherein the audio signal is included in the audio information bit string. Spectrum envelope information,
It separates and decodes voiced / unvoiced discrimination information, pitch period information, and gain information, obtains a spectrum envelope amplitude from the spectrum envelope information, and maximizes the value of the spectrum envelope amplitude among the bands divided on the frequency axis. Determining a band, based on the determined band and the voiced / unvoiced identification information, determining a mixing ratio when mixing a pitch pulse generated based on the pitch period information with white noise for each band; After creating a mixed signal for each band based on the determined mixing ratio, a mixed sound source signal is created by adding the mixed signals of all bands, and the spectrum envelope information and the gain information are obtained for the mixed sound source signal. To generate a reproduced sound. 2. A speech information bit sequence obtained by encoding a speech signal by a speech encoder based on a linear prediction analysis / synthesis method, comprising spectrum envelope information, voiced / unvoiced identification information in a low frequency band, and voiced in a high frequency band. A voice decoding method for reproducing a voice signal from a voice information bit sequence in which / unvoiced identification information, pitch period information, and gain information are encoded, wherein spectrum envelope information included in the voice information bit sequence,
Separating and decoding voiced / unvoiced identification information in a low frequency band, voiced / unvoiced identification information in a high frequency band, pitch period information and gain information, and based on the voiced / unvoiced identification information in the low frequency band, Determine the mixing ratio when mixing the pitch pulse and white noise generated based on the to create a mixed signal in the low frequency band, determine the spectral envelope amplitude from the spectral envelope information, the high frequency band on the frequency axis Of the divided bands,
Determining a band in which the value of the spectrum envelope amplitude is maximum; and determining the pitch pulse for each band obtained by dividing the high frequency band based on the determined band and the voiced / unvoiced identification information of the high frequency band. After determining the mixing ratio when mixing the white noise and creating a mixed signal, a mixed signal of the high frequency band is created by adding the mixed signals of all the bands of the high frequency band, A mixed sound source signal is created by adding the mixed signal of the high frequency band and the mixed signal of the high frequency band, and a reproduced sound is generated by adding the spectrum envelope information and the gain information to the mixed sound source signal. Audio decoding method. 3. A speech information bit sequence obtained by encoding a speech signal by a speech encoder based on a linear prediction analysis / synthesis method, comprising: spectrum envelope information, voiced / unvoiced identification information in a low frequency band, and voiced in a high frequency band. An audio decoding method for reproducing an audio signal from an audio information bit sequence in which / unvoiced identification information, pitch period information, and gain information are encoded, comprising: a spectrum envelope information included in the audio information bit sequence;
Separates and decodes voiced / unvoiced identification information in the low frequency band, voiced / unvoiced identification information in the high frequency band, pitch period information and gain information, and synchronizes with the pitch period based on the voiced / unvoiced identification information in the low frequency band. Determine the mixing ratio when mixing the white noise and the pitch pulse generated based on the pitch period information linearly interpolated, determine the spectral envelope amplitude from the spectral envelope information, high frequency band on the frequency axis Of the divided bands,
A band in which the value of the spectrum envelope amplitude is maximum is determined. Based on the determined band and the voiced / unvoiced identification information of the high frequency band, the pitch pulse and the pitch pulse are divided for each band obtained by dividing the high frequency band. The mixing ratio when mixing white noise is determined, and the spectrum envelope information, the pitch period information, the gain information, the mixing ratio of the low frequency band and the mixing ratio of each band obtained by dividing the high frequency band are pitch periods. Linearly interpolating in synchronization with the above, using the mixing ratio of the interpolated low frequency band, mixing the pitch pulse and white noise to create a mixed signal of a low frequency band, and interpolating the interpolated high frequency band. Using the mixing ratio of each divided band, the pitch pulse and the white noise are mixed to create a mixed signal for each divided band of the high frequency band, and then all the bands of the high frequency band To generate a mixed signal in the high frequency band, and to add the mixed signal in the low frequency band and the mixed signal in the high frequency band to create a mixed sound source signal, and to interpolate the mixed sound source signal. A speech decoding method, characterized in that a reproduced speech is generated by adding the subsequent spectral envelope information and the interpolated gain information. 4. The high frequency band is divided into three bands, and when the voiced / unvoiced identification information of the high frequency band indicates voiced, a mixing ratio of each band obtained by dividing the high frequency band is: When the value of the spectral envelope amplitude of the lowest frequency band or the second lowest frequency band is the maximum, the ratio of the pitch pulse is monotonously decreased as the frequency of the divided band becomes higher, and the highest frequency band is When the value of the spectral envelope amplitude is the maximum, the ratio of the pitch pulse in the band with the second lowest frequency is made smaller than the ratio of the pitch pulse in the band with the lowest frequency, and the ratio of the pitch pulse in the band with the highest frequency is the second. 4. The speech decoding method according to claim 2, wherein a predetermined amount is increased from a low frequency band. 5. The high frequency band is divided into three bands, and when the voiced / unvoiced identification information of the high frequency band indicates voiced, the mixing ratio of each band obtained by dividing the high frequency band is: The voiced intensity of the band where the value of the spectral envelope amplitude is maximum among the three bands is set to be larger than the voiced intensity of the same band when the value of the spectral envelope amplitude of the other bands is maximum. 3. The method according to claim 2, wherein
Alternatively, the audio decoding method according to 3. 6. The high frequency band is divided into three bands, and when the voiced / unvoiced identification information of the high frequency band indicates unvoiced, the mixing ratio of each band obtained by dividing the high frequency band is: The voiced intensity of the band where the value of the spectral envelope amplitude is maximum among the three bands is set to be smaller than the voiced intensity of the same band when the value of the spectral envelope amplitude of the other band is maximum. 3. The method according to claim 2, wherein
Alternatively, the audio decoding method according to 3. 7. A framer which inputs a sampled and quantized voice sample at a predetermined sample frequency and outputs a predetermined number of voice samples for each voice-encoded frame having a predetermined time length; RM which is the level information of the audio sample for one frame
A gain calculator that calculates the logarithm of the S value and outputs a logarithmic RMS value as a result thereof; and a first quantum that linearly quantizes the logarithmic RMS value and outputs the resulting logarithmic RMS value after the quantization. A linear prediction analyzer that performs a linear prediction analysis on the audio samples for one frame and outputs a linear prediction coefficient of a predetermined order that is spectrum envelope information; and an LSF (Line Spectrum Frequencie
s) an LSF coefficient calculator that converts and outputs the coefficients, a second quantizer that quantizes the LSF coefficients, and outputs an LSF parameter index that is a result thereof, A low-pass filter that filters at a cutoff frequency and outputs a low-frequency component of the input signal; and extracts a pitch period from the low-frequency component of the input signal based on a normalized autocorrelation function calculation, and calculates a pitch period and a normalized autocorrelation function. A pitch detector that outputs a maximum value, a third quantizer that performs a logarithmic transformation of the pitch period, linearly quantizes the pitch period at a first predetermined number of levels, and outputs a pitch period index that is a result thereof, Enter the maximum value of the normalized autocorrelation function, and set the aperiodic flag to ON if it is smaller than a predetermined threshold, and set it to OFF otherwise. An aperiodic flag generator that outputs the aperiodic flag; and an LPC analysis filter that removes spectral envelope information from the one-frame audio sample using the linear prediction coefficient as a coefficient and outputs a resulting residual signal. And a peakiness calculator that receives the residual signal, calculates and outputs a peakiness value, and corrects the value of the maximum value of the normalized autocorrelation function by the value of the peakiness calculator. A correlation function corrector that outputs the maximum value of the normalized autocorrelation function, and if the corrected maximum value of the normalized autocorrelation function is equal to or less than a predetermined threshold, it is determined that the voice is unvoiced; A first voiced / unvoiced determiner that outputs a voiced / unvoiced flag, and a non-uniformity of the pitch period of the frame whose aperiodic flag indicates aperiod at a second predetermined number of levels. An aperiodic pitch index generator for quantizing and outputting an aperiodic pitch index; inputting the voiced / unvoiced flag, the aperiodic flag, the pitch period index, and the aperiodic pitch index, and A periodic / aperiodic pitch / voiced / unvoiced information code generator for outputting a periodic / aperiodic pitch / voiced / unvoiced information code coded by the number of bits, and filtering the voice sample for one frame with a predetermined cutoff frequency A high-pass filter that outputs a high-frequency component of the input signal; a correlation function calculator that calculates and outputs a normalized autocorrelation function at a delay time given by the pitch period from the high-frequency component of the input signal; If the maximum value of the generalized autocorrelation function is equal to or smaller than a predetermined threshold, A second voiced / unvoiced decision unit that outputs a high-frequency voiced / unvoiced flag as a result thereof; a logarithmic RMS value after the quantization, the LSF parameter index, the periodic / aperiodic pitch voiced / unvoiced information code. And a bit packing unit that inputs the high-frequency voiced / unvoiced flag, performs bit packing for each frame, and outputs a voice information bit sequence. 8. An audio decoding device for decoding an audio information bit sequence for each frame encoded by the audio encoding device according to claim 7, wherein the audio information bit sequence is separated for each parameter, A non-periodic pitch / voiced / unvoiced information code, a quantized logarithmic RMS value, an LSF parameter index and a high-frequency voiced / unvoiced flag, and a bit separator for outputting the periodic / aperiodic pitch / voiced / unvoiced information code. If the state of the current frame is unvoiced, the pitch period is set to a predetermined value, and the voiced / unvoiced flag is set to 0 and output. A voiced / unvoiced information / pitch-period decoder for decoding and outputting based on rules, and setting and outputting a voiced / unvoiced flag to 1; When the current frame indicates unvoiced or non-periodic, the jitter value is set and output, and when the current frame indicates periodic,
A jitter setting unit that sets and outputs a jitter value to 0; an LSF decoder that decodes and outputs the LSF coefficient of the predetermined order from the LSF parameter index; and a slope that calculates and outputs a slope correction coefficient from the LSF coefficient. A correction coefficient calculator, a gain decoder that decodes the logarithmic RMS value after quantization and outputs gain information, and a first linear prediction coefficient calculator that converts and outputs the LSF coefficients to linear prediction coefficients. A spectrum envelope amplitude calculator that calculates and outputs a spectrum envelope amplitude from the linear prediction coefficient; a voice / unvoiced flag, the high-frequency voiced / unvoiced flag, and the spectrum envelope amplitude that are input and divided on a frequency axis. (Hereinafter referred to as “subband”) a pulse sound source / pulse sound source for determining and outputting mixing ratio information of a pulse sound source and a noise sound source.
A noise source mixing ratio calculator, linearly interpolating the pitch period, the mixing ratio information, the jitter value, the LSF coefficient, the tilt correction coefficient, and the gain information in synchronization with the pitch period, and the pitch period after interpolation A parameter interpolator for outputting the mixture ratio information after interpolation, the jitter value after interpolation, the LSF coefficient after interpolation, the slope correction coefficient after interpolation, and the gain information after interpolation, and a pitch period after interpolation and a parameter after interpolation. A pitch period calculator that inputs a jitter value, adds the jitter to the pitch period after interpolation, and outputs a pitch period converted into an integer value (hereinafter referred to as an “integer pitch period”); A one-pitch waveform decoder for synchronously decoding and outputting the reproduced voice for the integer pitch period, the one-pitch waveform decoder comprising a single pulse signal within the integer pitch period period A single pulse generator for outputting, a noise generator for outputting white noise having a length of the integer pitch period, and a single pulse signal for each subband based on the interpolated mixing ratio information. A mixed sound source generator that mixes them with white noise and then combines them to output a mixed sound source signal; a second linear prediction coefficient calculator that calculates a linear prediction coefficient from the interpolated LSF coefficients; This is a cascade connection of an adaptive pole / zero filter having coefficients obtained by performing bandwidth expansion processing on prediction coefficients and a spectrum inclination correction filter having coefficients of the interpolation-corrected interpolation coefficients. An adaptive spectrum enhancement filter that outputs a sound source signal having an improved spectrum; and an all-pole filter that uses the linear prediction coefficient as a coefficient. An LPC synthesis filter that adds spectrum envelope information to the generated sound source signal and outputs a signal to which the spectrum envelope information has been added; and a gain using the gain information for the signal to which the spectrum envelope information has been added. An audio decoding device comprising: a gain adjuster that performs adjustment and outputs a reproduced audio signal; and a pulse spreading filter that performs pulse spreading processing on the reproduced audio signal and outputs a reproduced audio signal that has been subjected to pulse spreading processing.