JP3404350B2 - Speech coding parameter acquisition method, speech 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 a speech coding parameter acquisition method and apparatus for digitizing a speech signal to obtain speech coding parameters representing its characteristics at predetermined time intervals, and the speech coding parameter. The present invention relates to a voice decoding method and apparatus for synthesizing an original voice signal based on, transmitting or storing voice encoding parameters by encoding, and restoring voice encoding parameters when necessary from a transmission destination or an accumulation destination, It is suitable for use in a digital mobile phone, a digital voice storage device, or the like that synthesizes a voice signal from a restored voice coding parameter and transmits a voice.

[0002]

2. Description of the Related Art Since a digitized voice signal can be subjected to various digital signal processing such as data compression, error processing, and multiplexing, it is not limited to fixed-line telephones and mobile telephones, but can be applied to multimedia systems using voice. Widely adopted. To digitize analog audio signals,
Generally, it is necessary to perform sampling at a sampling frequency that is at least twice the input voice frequency band and perform quantization with a quantization step that cannot be discriminated by the ear, and thus requires a wider transmission frequency bandwidth than an analog signal. . Therefore, the audio signal once digitized is compressed by various encoding and modulation methods according to the required audio quality. By actively utilizing the characteristics of voice, efficient compression can be performed. For example, the adaptive differential pulse code modulation (ADPCM) method is a waveform coding method that uses the logarithmic characteristics of the speech waveform periodicity and human auditory sensitivity. It compresses 128 kbps digital speech to about 32 kbps, and is the same as before compression. It has no voice quality and is used for backbone transmission of telephones and PHS systems. In the waveform coding method, since the sampling point is represented by at least 1 bit, it is theoretically impossible to reduce the speech coding speed to 8 kbps or less when the sampling frequency is 8 KHz.

In order to obtain a low speech coding rate, speech is divided into segments at predetermined time intervals, and code excitation linear prediction (CELP) is basically used for transmitting speech synthesis parameters and residual excitation signals for each segment. There is a method.
VSELP and PS used in Japanese mobile phones
In the I-CELP method, a linear prediction coefficient (LPC) that approximates human vocal tract filter characteristics obtained by linear prediction analysis of a voice signal at intervals of 20 msec or 40 msec and a waveform that is audibly similar to the input voice can be synthesized. A low speech coding speed is realized by coding the residual error source signal. Moreover, in order to efficiently encode the residual excitation signal, a codebook having a plurality of residual excitation waveforms is prepared, and the entry number and gain of the codebook are transmitted. These details are described in detail in the standard RCR-STD27F of the Association of Radio Industries and Businesses. This CELP-based method is designed to properly design a codebook of an appropriate size to achieve a voice coding rate of 3 to 4 kbps.
It has been realized to a degree.

In order to obtain a still lower voice coding speed, there is a system in which only the voice synthesis parameters are transmitted to perform voice coding without using the excitation codebook in the CELP system. US Department of Defense standard speech coding scheme FS-1015
Is the pitch frequency, LPC coefficient, root mean square amplitude,
A voice coding method of an LPC vocoder system that performs voice coding / decoding based on a voice synthesis parameter of voiced / unvoiced determination information has achieved a voice coding rate of 2.4 kbps. This method actively uses the characteristics of voice,
It has a drawback that the quality of the synthesized speech becomes like that of the synthesized speech, and the quality of the decoded speech remarkably deteriorates especially in the background noise. In addition, IMBE (Improved Multiba), which is partially used in satellite mobile phones
nd Excitation) method is a method of converting a voice time segment into a frequency domain and performing voice encoding with voice pitch / voice harmonics amplitude and voiced / unvoiced information of a frequency band obtained by dividing a frequency band into a plurality of voice segments. Since a voiced sound model and an unvoiced sound model are selected and synthesized for each band, it is reported that the synthesized speech is less deteriorated even in the presence of background noise and mixed speech, and is superior to the LPC vocoder. .

FIG. 14 is a diagram showing the configuration of a general voice coding transmission device. Speech coding parameter extraction unit 3
Reference numeral 02 divides the sampled and quantized voice digital signal input from the voice input terminal 301 into segments at predetermined time intervals, and extracts voice coding parameters for each segment. The voice encoding parameters to be extracted are determined by the voice encoding method. For example, in the IMBE method, the voice pitch, the amplitude of voice harmonics, and the voiced / unvoiced information of each frequency band are used. Parameter coding unit 30
3 effectively encodes the extracted speech coding parameter to reduce the code amount, and sends it to the transmission path 305 via the transmitting unit 304. The parameter decoding unit 307 decodes the code received by the receiving unit 306 to restore the voice coding parameters, and the voice synthesizing unit 308 creates synthetic voice by the operation reverse to the operation of the voice encoding parameter extracting unit and outputs the voice. An audio digital signal is output from the terminal 309.

FIG. 15 is an internal configuration diagram of the speech coding parameter extraction unit 302 in the case of the IMBE method. The digital input voice signal 301 is input to the fundamental frequency estimation unit 401, where the fundamental frequency of the voice is estimated. There are many methods for estimating the fundamental frequency, such as a method for detecting a cepstrum peak, which is an inverse Fourier transform of an autocorrelation function or a logarithm of a frequency spectrum. For example, Furui “Digital Speech Processing” Tokai University Press, 1985. It is described on September 25, etc. The frequency spectrum calculation unit 402 frequency-analyzes a finite-length voice segment cut out by a window function such as a Hamming window to obtain a voice frequency spectrum. The fundamental frequency correction unit 403 detects Ab-S (Analysis-by-Synthesi) in a frequency range near the speech fundamental frequency estimated by the fundamental frequency estimation unit 401.
s) The fundamental frequency ωo corrected by the minimum error condition between the synthesized spectrum and the voice frequency spectrum calculated by the frequency spectrum calculation unit 402 is obtained by the method. The voiced strength calculation unit 404 divides the frequency band into a plurality of frequency bands (frequency sections) k based on the corrected fundamental frequency ωo.
Divide into (k = 1,2, ..., K), calculate the error between the synthesized spectrum synthesized for each frequency band and the voice frequency spectrum, and output the voiced / unvoiced information V [k] by threshold judgment To do. The spectrum envelope calculation unit 405 uses the voiced / unvoiced information V
From [k], the root mean square value (RMS value) of the frequency spectrum of the frequency spectrum in the frequency band of each harmonics obtained by the A-B-S method in the voiced band, and in the unvoiced band, the spectrum envelope absolute value | Output as A (ω) |.

FIG. 16 is a diagram showing an internal configuration of the voice synthesis unit 308 in the case of the IMBE method. As shown in this figure, the voice synthesis unit 308 is roughly divided into a voiced voice synthesis unit 508 and an unvoiced voice synthesis unit 509. In the voiced voice synthesis unit 508, the harmonic sound source unit 501 uses the voiced / unvoiced information V [k] and the fundamental frequency ωo in the frequency section determined to be voiced to generate a spectral envelope | A plurality of sound source signals are generated by driving with an amplitude corresponding to A (ω) |. Harmonic adder 502
Then, a plurality of sound source signals generated in the harmonic sound source unit 501 are added and combined to generate a voice signal corresponding to the voiced band.
Further, in the unvoiced speech synthesis unit 509, the noise source unit 50
In No. 3, white noise is generated, processed by an appropriate window function in the frequency conversion unit 504, and then Fourier transformed (FFT) to be converted into a frequency domain signal. In the noise extraction unit 505, the white noise spectrum of the frequency band designated as unvoiced by V [k] is extracted from the white noise converted into the frequency domain signal, and the power for each frequency band of the spectrum envelope | A (ω) | Adjust the amplitude of each spectrum so that
The inverse frequency conversion unit 506 converts the frequency spectrum of the noise section corresponding to the unvoiced band into an inverse Fourier transform (IFFT) to convert it into a speech waveform. In the addition unit 507,
The voiced speech waveform from the harmonic addition section 502 of the voiced speech synthesis section 508 and the inverse frequency conversion section 5 of the unvoiced speech synthesis section 509.
The unvoiced voice waveform converted into the time-axis waveform signal in 06 is added to obtain the final synthesized voice having voiced sound and unvoiced sound. For details of this IMBE method, see "Multiband Excitati
on Vocoder ”, IEEETransactions on Acoustics, speec
h, and signal processing, vol.36, No.8, August 1988,
It is described in detail in pp1223-1235. As described above, as a method for digitizing voice to realize low-bit-rate voice encoding, an analysis-synthesis type voice encoding such as IMBE method for extracting and encoding voice encoding parameters based on a voice synthesis model A scheme has been proposed.

[0008]

As described above, the analysis-synthesis type speech encoding method is effective for low bit rate speech encoding, but it does not use the residual excitation signal and uses the speech synthesis parameter. Since only the voice is synthesized, the synthetic voice quality tends to be obtained depending on the encoding method. The IMBE method, which is an analysis-synthesis type audio encoding that realizes low bit rate audio encoding, divides the input audio into segments, cuts out an audio frame, divides the frequency band of the frame into a plurality of bands, and divides the band into the bands. Determine whether the included frequency component is voiced or unvoiced, set a voiced sound synthesis model and an unvoiced sound synthesis model for each band, and add them to obtain a synthesized voice. It improves the synthetic speech quality in mixed voiced and unvoiced frames. The accuracy of estimation of the voice fundamental frequency (voice pitch), voiced / unvoiced information, and spectral envelope information as the encoding parameters is important in determining the quality of the reproduced voice. Although the voice fundamental frequency can be obtained by the above-mentioned autocorrelation method, the IMBE method shows a method in which the fundamental frequency is estimated with 1/2 pitch accuracy by evaluating with an autocorrelation function extended to an integral multiple pitch. . Further, in order to obtain the spectrum envelope, the extracted fundamental frequency ωo and the frequency spectrum of the frequency analysis window are used for estimation by the above-mentioned A-B-S method. However, the spectrum envelope is estimated with the accuracy of the estimated fundamental frequency. Since the accuracy of the estimation is insufficient, a method of simultaneously estimating the spectrum envelope while searching the vicinity of the estimated fundamental frequency with ¼ pitch accuracy is adopted.

The procedure for obtaining this spectrum envelope is as follows. First, the signal s (n) of the input speech segment
Is limited to −110 to 110 samples by the frequency analysis window w _R (n), and then the frequency spectrum S _w (m) is obtained by the formula (1) by the FFT of 256 stages.

[Equation 1] Next, the Lth harmonic of the fundamental frequency ωo (L = 1,2, ..., Lma
x; (Lmax + 0.5) · ωo <2π) with the center frequency as the center frequency and the spectrum of the frequency analysis window w _R (n) E _w (ω), and the envelope value A _L of the equation (2) Then, the spectrum S _w (m) of the voice segment is approximated by the sum of the above, and the envelope value A _L of each harmonic is obtained.

[Equation 2]

At this time, the envelope value A _L of each harmonic is calculated by the error least squares method while changing ω o with 1/4 pitch accuracy, and the spectrum error evaluation value E is calculated from the obtained ω o and A _L.
(ωo)

[Equation 3] Is the fundamental frequency correction value, and the amplitude A _L of each harmonic at that time is the spectrum envelope value. Also,
The voiced strength of each frequency band (al to bl) is estimated by determining the spectral error relative value Dk shown in the equation (4) as a threshold value.

[Equation 4]

Here, the changing step when searching for ωo and the effect of the error of the actual voice fundamental frequency on the determination result will be considered. The fundamental frequency of voice varies between individuals and men and women, but the central frequency of men is approximately 125 Hz, and the fundamental frequency of women is approximately double, with total frequencies of 70 Hz to 400 Hz.
It is in the Hz range. The error of the fundamental frequency to be evaluated is expanded to L times the frequency error of the L-th harmonic. Table 1 shows the frequency error Δf generated by the pitch error ΔP of the voice fundamental frequency ωo (= 2πfo) and the frequency error Δf (2 kHz) in the harmonic region near 2 kHz calculated by the equation (5). Where fs is the sampling frequency of the voice segment.

[Equation 5]

[0012]

[Table 1]

As can be seen from Table 1, when the estimation error ΔP of the actual voice fundamental frequency (voice pitch) is 1 pitch, the frequency error of the harmonic near 2 kHz is ± 25 to ± 75H.
Up to z, the spectrum interval becomes 8000/256 = 31.25 Hz or more when frequency analysis is performed by 256-stage FFT. Also, f
At o = 300 Hz, the harmonic spectrum error near 2 kHz becomes 38 Hz when ΔP is 0.5 pitch, and finally becomes 19 Hz when ΔP is 0.25. On the other hand, FIG. 17 shows the spectrum evaluation error in the case of FFT 256 stages calculated by the equation (4) using the frequency analysis window as the Hamming window. For example, in the case of a standard woman having a fundamental frequency of fo = 275 Hz, the fundamental frequency pitch is Pi = 29, but if the estimation of the fundamental frequency is estimated as Pi = 28 and there is an error of -1 pitch, The estimated fundamental frequency is fo = 8000 / (29-1) = 285.7 (Hz), the fundamental frequency error is Δfo = 10.7 Hz, and the normalized spectrum error Dk = 0.1 is obtained from FIG. Furthermore, in the case of the estimated pitch Pi = 27, the estimation error of -2 pitch and the fundamental frequency error are 2
1 Hz and Dk = 0.3, which has a large effect on the voiced / unvoiced determination due to the normalized spectral error. Further, in the case of harmonics harmonics, the fundamental frequency error is magnified by the harmonic order. For example, in the case of 1/4 pitch error,
The estimated fundamental frequency is 8000 / (29-0.25) = 278.26 (Hz), and the fundamental frequency error Δfo = 3.26Hz, but in the vicinity of 2kHz it is 2000/275 times expanded to a frequency error of 23.7Hz,
From FIG. 17, the normalized spectrum error expands from about 0.01 to 0.35 or more, which causes a voiced / unvoiced decision error. The voiced / unvoiced information and the spectral envelope information are also parameters that characterize the entire speech segment, and errors in these estimations have a large influence on the quality of coded speech, as already mentioned.

Further, in the voice decoding in the IMBE system, as shown in FIG. 16, a random noise source is frequency-converted (FFT) for unvoiced sound, and only the frequency range of unvoiced sound designated by the voice coding parameter is extracted. Then, inverse frequency conversion (IFFT) is performed to synthesize unvoiced voice. In this case, two steps of frequency conversion are required,
In particular, there is a drawback that the calculation load is large when the update period of the voice segment is set to be short in order to improve the encoded voice quality.

Therefore, it is an object of the present invention to provide a speech coding method and apparatus capable of performing highly accurate voiced strength determination regardless of changes in the fundamental frequency of speech and having strong error resistance against spectrum noise. I am trying. Another object of the present invention is to provide a speech decoding method and device with a small calculation load.

[0016]

In order to achieve the above object, a method for acquiring a voice coding parameter according to the present invention is a method for converting a digitized voice signal into a predetermined repeating cycle,
A method for acquiring a voice coding parameter for acquiring a voice coding parameter from a voice segment extracted in a predetermined segment length, the step of acquiring a voice fundamental frequency from the voice segment, a variable length determined by the voice fundamental frequency Acquiring a first frequency spectrum from a variable length segment from which the audio signal is extracted by an adaptive window of, a step of acquiring a second frequency spectrum from a fixed length segment from which the audio signal is extracted by a fixed length window, Dividing the first frequency spectrum into a plurality of frequency bands, frequency spectrum power of the first frequency spectrum, frequency spectrum power of each frequency band, number of harmonics included in each frequency band, harmonics amplitude of each harmonics. And Harmo Determining a voiced strength for each of the respective frequency bands by box bandwidth, and the second
And calculating the spectral power of each harmonics band divided so that the frequency bandwidth becomes the voice fundamental frequency with a frequency that is an integral multiple of the voice fundamental frequency as the center. The length of the variable-length adaptive window is determined by the relationship between the bandwidth of the frequency spectrum distribution of the variable-length adaptive window and the voice fundamental frequency. Further, the variable-length adaptive window is a Hamming window having a length of four times or more the cycle corresponding to the voice fundamental frequency.

Furthermore, in the voice decoding method of the present invention, the voice fundamental frequency of the voice segment obtained by extracting the digitized voice signal at a certain fixed cycle and the frequency spectrum of the voice segment are integers of the voice fundamental frequency. The spectrum power of each harmonics band divided so that its frequency bandwidth becomes the voice fundamental frequency, and the frequency band obtained by dividing the frequency spectrum of the voice segment into a plurality of frequency bands is voiced or unvoiced. A voice decoding method for synthesizing voice by a voice encoding parameter consisting of discriminated information, wherein the discriminant information has a frequency band in which a voice is present, the center frequency of which has an integral multiple of the voice fundamental frequency. In addition, the amplitude that is equal to the spectral power of the corresponding harmonic band is In the frequency band in which the discrimination information indicates unvoiced, the central symmetric random sequence and the central antisymmetric random sequence are regarded as the real part and the imaginary part of the frequency spectrum sequence of the noise signal, and An interval corresponding to the frequency band is extracted from one random sequence, amplitude adjustment is performed so as to be the same as the spectral power of the corresponding harmonics band, and the real part thereof is obtained by inverse Fourier transform to obtain an unvoiced frame signal. An unvoiced voice is generated by linearly interpolating between the unvoiced frame signal of the previous segment and the unvoiced frame signal obtained this time, and then added to the generated sine wave group to obtain a synthesized voice.

Furthermore, the speech coding parameter acquisition apparatus of the present invention acquires speech coding parameters from a speech segment obtained by extracting a digitized speech signal at a predetermined repetition period with a predetermined segment length. An apparatus for acquiring a voice coding parameter, comprising means for acquiring a voice fundamental frequency from the voice segment, and a first variable length segment obtained by extracting the voice signal by an adaptive window of a variable length determined by the voice fundamental frequency. Means for obtaining a frequency spectrum, means for obtaining a second frequency spectrum by a fixed-length segment obtained by extracting the audio signal through a window of a fixed length, means for dividing the first frequency spectrum into a plurality of frequency bands, and 1 frequency spectrum to frequency spectrum power, frequency band of each said Means for determining the voiced strength for each frequency band based on the number spectrum power, the number of harmonics included in each frequency band, the harmonics amplitude of each harmonics and the harmonics bandwidth, and the voice fundamental frequency from the second frequency spectrum It has means for calculating the spectral power of each harmonics band divided so that its frequency bandwidth becomes the fundamental frequency of the voice with the frequency being an integral multiple of.

Furthermore, in the speech decoding apparatus of the present invention, a speech fundamental frequency of a speech segment obtained by extracting a digitized speech signal at a certain repetition period and a frequency spectrum of the speech segment are integers of the fundamental speech frequency. The spectrum power of each harmonics band divided so that its frequency bandwidth becomes the voice fundamental frequency, and the frequency band obtained by dividing the frequency spectrum of the voice segment into a plurality of frequency bands is voiced or unvoiced. A voice decoding device for synthesizing voice by a voice encoding parameter consisting of discriminated discrimination information, wherein the discrimination information has a frequency band indicating voiced voice, the center frequency of which has an integral multiple of the fundamental frequency of the voice, In addition, the amplitude that is equal to the spectral power of the corresponding harmonic band is Means for generating a sine wave group, a means for generating a noise signal of a central symmetric random sequence and a central antisymmetric random sequence, and a section corresponding to the frequency band in which the discrimination information is unvoiced, from the two random sequences. Means for adjusting the amplitude of the extracted random-sequence noise signal so that its spectral power becomes the same as the spectral power of the harmonics band corresponding to the frequency band in which the discrimination information indicates unvoiced, and the amplitude is adjusted. Means for inverse Fourier transforming a random sequence noise signal to generate an unvoiced frame signal, means for linearly interpolating the unvoiced frame signal of the preceding segment and the unvoiced frame signal this time, and an unvoiced speech signal, and It has means for adding the generated sine wave group and the generated unvoiced voice.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION The speech coding parameter acquisition method, speech decoding method and apparatus according to the present invention stably estimate speech coding parameters in speech coding, particularly in low bit rate speech coding. The speech coding parameter extracting unit 302 of the speech coding transmission apparatus shown in FIG. 14 can be incorporated and used in the method and apparatus, and the method and apparatus for speech decoding by the estimated speech coding parameter. , And speech synthesizer 308
The case where the present invention is applied to will be described as an example. Also,
The present invention can be applied to various speech coding systems, but here, the case of being applied to the IMBE system will be described as an example.

FIG. 1 is a block diagram of a speech coding parameter extraction section to which the speech coding parameter acquisition method of the present invention is applied. As shown in this figure, the speech coding parameter extraction unit of the present invention includes a fundamental frequency estimation unit 401 that estimates the fundamental frequency ωo of the input speech signal 301.
A voiced strength calculation unit 404 that divides a frequency spectrum obtained by frequency-analyzing the input voice signal frame into a plurality of frequency bands, and outputs voiced strength information V [k] indicating voiced / unvoiced for each band, and The input voice signal is frequency-analyzed using a fixed-length window, and is composed of three parts of a spectrum envelope calculation unit 405 that calculates a spectrum envelope | B (ω) |.

In the speech coding parameter acquisition method of the present invention, the frequency spectrum error between the synthesized speech and the input speech is not used as a voiced / unvoiced evaluation value as in the conventional method, but the input speech is evaluated. The harmonics amplitude of the voice included in a certain frequency band of the frequency spectrum is measured from the input voice spectrum amplitude, and the harmonics amplitude is used as an evaluation value for voiced strength or voiced / unvoiced judgment. Then, in measuring the frequency spectrum of the input voice, the frequency resolution is adjusted by adaptively adjusting the width of the spectrum analysis window according to the estimated fundamental frequency of the input voice, without unnecessarily decreasing the time resolution. , We are taking the necessary frequency resolution. In addition, the number of harmonics included in each frequency band is measured, and the number of harmonics is used as another evaluation value to determine how close it is to the expected number of harmonics. Furthermore, the certainty of the determination is improved by measuring the frequency width of each harmonic (harmonic width) and determining how close it is to the expected harmonic width by the spectrum analysis window. Furthermore, based on the knowledge that the input voice is silent when the power (energy) is small, the frequency spectrum power of the input voice and the voice frequency spectrum power of each frequency band are added to the evaluation value. Also, in extracting the spectrum envelope,
A frequency spectrum of the input voice is extracted by a second frequency analysis using the spectrum analysis window and a fixed-length window having a different analysis length, and the frequency spectrum of the input voice is in the region of the voice fundamental frequency width at every frequency interval of an integral multiple of the estimated voice fundamental frequency. It is extracted as the square root of the spectral power.

The reason for this will be further described with reference to FIGS. 3, 4 and 5. FIG. 3 shows an example of the frequency spectrum amplitude value (logarithmic value) of a voice segment made up of almost voiced voice. The horizontal axis represents the discrete frequency when the fast discrete Fourier transform (FFT) of 256 points is performed. As shown in this figure, a clear harmonic spectrum with a certain width is observed at a certain interval in the spectrum amplitude,
Its logarithmic amplitude and width also have stable amplitude over a wide range of frequencies. From this, it can be expected that the harmonics amplitude and its number within a certain frequency band can be measured without being affected by the estimation error of the fundamental frequency ωo. Also, FIG.
Is an example of the frequency spectrum amplitude value (logarithmic value) of a voice segment with many unvoiced voices. In this case, it can be seen that the harmonics amplitude and the width of the harmonics within the defined frequency band are small, and the number of harmonics above a certain level is also small, and that value is influenced by the estimation error Δωo of the fundamental frequency ωo. You can read that you do not receive much. Based on the above consideration, in the voiced / unvoiced determination, the harmonics amplitude measured from the frequency spectrum amplitude logarithmic value, the number of effective harmonics with an amplitude above a certain threshold, the width of the harmonics, the power of the input voice, and the frequency band The voice power is used for judgment and evaluation.

The length of the frequency analysis window (time range) T
Considering the relationship between the _w (sec) or L _w (sample) and the voice fundamental frequency fo (Hz) or the pitch P (sample), as shown in FIG. 5, the harmonic center appears in the spectrum amplitude at an integral multiple of the fundamental frequency. When the frequency analysis window is the Hamming window, the bandwidth of each harmonic is 4 / T _w . Therefore, the length of the frequency analysis window is determined by the equation (6) on condition that the valley of harmonics does not enter the center from the valley of adjacent harmonics.

[Equation 6]

The length of the frequency analysis window may be changed to be four times the basic pitch in accordance with the equation (6), and may be changed in proportion to the basic frequency. You may set it. For example, switching may be performed every time the pitch increases by 20, and the first frequency analysis window may be set according to the change in pitch according to the equation (7).

[Equation 7] Here, ceil (x) is a function that gives the smallest integer exceeding x, and the analysis window length is an odd number with central symmetry and prevents the analysis window length from becoming a length other than that expected from the pitch range. Because of that there are restrictions. FIG. 6 shows the extracted basic pitch P.
An example of setting L _w with respect to Further, the window function value w when the adaptive Hamming window shown in Expression (8) is used as the window function
The calculation result of _s (n) is shown in FIG. (However, the number of FFT stages M =
Calculated as 512)

[Equation 8]

Further, in the present invention, the FFT required depending on how many spectrums p are set between the fundamental frequencies fo.
You can decide the number of steps M. For example, the sampling frequency is 8
In the case of kHz, it can be determined by the equation (9).

[Equation 9] If the minimum fundamental frequency is 60Hz, p =
In case of 4, M = 533, and FFT of about 512 stages
You can see that the number of steps is required.

As described above, according to the speech coding parameter acquisition method of the present invention, the frequency analysis window in which the window size is adaptively set by the speech fundamental frequency (pitch) is used,
Since the frequency spectrum is obtained by the FFT having the number of stages determined from the range of the fundamental frequency, it is possible to reduce the mutual interference of spectrum between harmonics. Then, from the frequency spectrum thus obtained, the harmonics amplitude, the number of harmonics, the harmonics width, the frame energy (frame power), and the band energy (band power) are measured to obtain voiced / unvoiced information for each frequency band. Therefore, the detailed pitch is unnecessary for voiced / unvoiced determination, and the possibility of a determination error due to a pitch error can be reduced. Further, the spectrum envelope information is obtained separately from the voiced / unvoiced determination, and is obtained from the square root of energy for each harmonic band by FFT with a fixed window size. Therefore, even if there is an error in the unvoiced / voiced determination, it is possible to obtain spectrum envelope information that is not affected by the error.

FIG. 2 is a block diagram showing an example of the configuration of a speech decoding apparatus to which the speech decoding method of the present invention is applied. As shown in this figure, the speech decoding apparatus has a fundamental frequency ωo and a spectral envelope | B of the decoded speech parameters.
(ω) | is input and the voiced voice synthesis unit 5 that synthesizes voiced voice
08, voiced / out of decoded speech coding parameters
The unvoiced information (voiced strength information) V [k] and the spectrum envelope | B (ω) | are input, and are composed of an unvoiced speech synthesis unit 509 and an addition unit 507. Here, the voiced voice synthesis unit 508 is similar to the conventional voice synthesis unit 508 shown in FIG.
The harmonic sound source unit 501 generates a harmonic signal corresponding to the fundamental frequency ωo and the voiced frequency band based on the fundamental frequency ωo and voiced / unvoiced information, and the spectrum Envelope information | B (ω) |
Based on, control the amplitude of the signal of each of those frequencies,
The harmonic addition unit 502 adds them.

The unvoiced speech synthesizer 509 includes a symmetric random sequence generator 201 and an antisymmetric random sequence generator 20.
2, random sequence extraction unit 203, inverse frequency conversion unit 204
And a frame interpolating unit 205. Then, in the frequency band in which the voiced / unvoiced discrimination information V [k] indicates unvoiced, the centrally symmetric random sequence and the antisymmetric random sequence generation unit 202 generated by the symmetric random sequence generation unit 201.
The central antisymmetric random sequence generated in step S1 is regarded as the real part and the imaginary part of the frequency spectrum sequence of the noise signal, and the corresponding random frequency band is extracted from the two random sequences in the random sequence extraction unit 203, and its power After the amplitude is adjusted to be the same as the power of the corresponding unvoiced harmonics band, the inverse frequency transform unit 204 performs an inverse Fourier transform (IFFT) to obtain its real part, which is used as an unvoiced frame signal. At 205, an unvoiced voice is generated by linearly interpolating between the previous unvoiced frame signal and the frame, and then added by the adder 507 with the generated sine wave group to obtain a synthesized voice.

That is, in the speech decoding method of the present invention, the frequency spectrum corresponding to the noise source is not created from the noise source by FFT as in the conventional method, but the symmetric random sequence generator 201 and the antisymmetric random sequence generator are generated. A method of directly creating a frequency spectrum corresponding to a noise source from the random noise sequence generated by the unit 202 is used. Then, the frequency band corresponding to the unvoiced frequency band is extracted from the frequency spectrum,
After generating unvoiced speech on the real-time axis by inverse FFT,
An unvoiced sound having a required frame length is synthesized by linear interpolation in which the interpolation weight becomes 1 between frames. As a result, unvoiced speech can be synthesized by only one inverse Fourier transform, and the amount of calculation can be reduced.

Here, it is necessary to set conditions for the generation of the random sequence. This is the same as the condition that the imaginary part does not occur and the total power of the FFT spectrum appears in the real time axis sequence when converted to the time axis sequence by the inverse FFT. This condition can be expressed by equation (10).

[Equation 10] Here, Sw (m) is a random sequence regarded as a frequency spectrum, Re is a real part, Im is an imaginary part, m is a frequency spectrum, and when m = 0 represents a DC component.

The speech coding apparatus to which the speech coding method of the present invention shown in FIG. 1 is applied will be described in more detail. In FIG. 1, an audio digital signal sampled at a sampling frequency of about 8 kHz, which is input from an audio input terminal 301, is input to a fundamental frequency estimation unit 401, and here, for example, a fixed length is set at every 20 msec time interval. A voice segment (frame) is taken out, and a voice fundamental frequency ωo in the segment is estimated. As described above, the method of estimating the fundamental frequency includes the method of using the autocorrelation and the method of using the cepstrum.

The voice digital signal is also input to the voiced strength calculation unit 404 and the spectrum envelope calculation unit 405. In the spectrum envelope calculation unit 405, the second
The spectrum calculation unit 111 calculates a discrete frequency spectrum value B [m] by performing a fast Fourier transform (FFT) on the signal obtained by windowing the speech segment with a window function such as a Hamming window in the fixed window processing unit 110. . When the sampling frequency of the digital audio input signal is fs and the FFT of 256 points is performed, the calculated frequency spectrum B [m] is calculated for each frequency interval fd represented by the following formula (11).

[Equation 11]

The spectrum power calculation unit 112 has the frequency spectrum B for each harmonic band having a bandwidth equal to the fundamental frequency centered on a frequency that is an integral multiple of the fundamental frequency ωo estimated by the fundamental frequency estimation unit 401. The square root of the sum of squares of [m] is calculated, and this is output as the spectrum envelope value | B (ω) |.

The voiced strength calculation unit 404 is adapted to the adaptive window processing unit 1.
01, the first spectrum calculation unit 102, the frame energy calculation unit 103, the band energy calculation unit 104, the logarithmic conversion unit 105, the band harmonics amplitude calculation unit 106,
The band harmonics width calculation unit 107, the band harmonics number calculation unit 108, and the voiced strength determination unit 109 are configured.

The adaptive window processing unit 101 adaptively sets the voice signal s (n) from the voice fundamental frequency ωo estimated by the fundamental frequency estimating unit 401 by the above equations (6) to (8). Window processing is performed using a Hamming window of length, and the first spectrum conversion unit 102 obtains the frequency spectrum A [m] of the voice segment by the FFT shown in Expression (12). Where M
Is the number of FFT samples.

[Equation 12]

The frame energy calculation unit 103 calculates from the frequency spectrum A [m] the average energy (referred to as "frame energy" or "frame power") E of the frame in which the amount of energy decrease due to the adaptive window is compensated.
f is calculated by the equation (13). Here, the second term compensates the energy decrease due to the window function.

[Equation 13]

The band energy calculation unit 104 calculates the average energy for each frequency band (referred to as "band energy" or "band power") Eb [k] (k = 1, ..., K).
The band energy Eb [k] is expressed by the following equation (14), where [ak, bk] is the spectrum section of the k-th band.

[Equation 14] Here, when the frequency range of the band is set to three times the fundamental frequency ωo, ak and bk are

[Equation 15] become. However, floor (x) indicates the maximum integer that does not exceed x.

The logarithmic conversion unit 105 has the frequency spectrum value | A [m] | calculated by the first spectrum calculation unit 102.
And the logarithmic spectrum amplitude sequence LA [m] is calculated.

[Equation 16]

The band harmonics amplitude calculator 106 is
Calculate the harmonics amplitude Ah or Bh in each frequency band. A method of evaluating the harmonics amplitude will be described with reference to FIG. Harmonics amplitude is the spectral amplitude |
It is the difference between the maximum value of the data string of A [m] | and the minimum value of its nearest neighbor, but when the harmonics amplitude is expressed linearly, the amplitude increases or decreases in proportion to the spectrum intensity. Therefore, the maximum value H0 of the spectrum amplitude and the minimum values H before and after it
If the values Ah1 and Ah2 obtained by normalizing the difference between 1 and H2 with the maximum value H0 are used as the evaluation values of the harmonic amplitude, the harmonic intensity not related to the spectral intensity can be evaluated. Where A
If the smaller of h1 and Ah2 is the harmonics amplitude evaluation value Ah,

[Equation 17] Becomes Alternatively, the harmonics evaluation value B, which represents the harmonics intensity by the ratio of the spectrum maximum value and the spectrum minimum value,
You may evaluate with h. That is,

[Equation 18] These Bh1 and Bh2 represent the amount of attenuation from the peak of the harmonics in decibel units, and from the result of the spectral amplitude measurement of the voice shown in FIG. You can see that it is an evaluation unit.

The band harmonics width calculation unit 107 receives the output of the logarithmic conversion unit 105, and determines the frequency interval between the local minimum value immediately before and the local minimum value immediately after each of the spectral amplitude maximum values as the width of the harmonics. calculate. The band harmonics number calculation unit 108 includes the logarithmic conversion unit 105.
The number H of harmonics included in the frequency spectrum range of the frequency band shown in the above equation (15)
Calculate n. The harmonics number is calculated by the frequency spectrum amplitude from discrete frequencies ak to bk obtained by FFT.
20log ₁₀ | A [m] | and the spectrum amplitude before and after it 20log ₁₀ |
A [m-1] |, 20log ₁₀ | A [m + 1] | are compared. If there is more than either value, the mth spectrum is the spectrum closest to the center frequency of harmonics at the maximum point of the spectrum. To judge. That is,

[Formula 19]

Here, if the calculated spectrum logarithmic value is used as it is to count the number of local maxima by the above method, there is a harmful effect of counting the spectral maxima due to noise due to the influence of spectral noise. Noise is removed to prevent erroneous counting due to noise. A method of removing the spectrum noise will be described with reference to FIG. 9A and 9C show the case where there is spectrum noise, and the spectrum amplitudes of m + 1 and m + 2 are reversed. The sign of the difference between the spectrum amplitudes for a set of four consecutive spectra is-+-or +-
In the case of +, there is a maximum value, and each maximum value is m +
It appears at the 2nd or m + 1st position, and the amplitude of the maximum value is m +
It can be seen that there is a difference between the first and m + 2th amplitudes. Therefore, if the difference between the m + 1-th and m + 2-th spectra is smaller than a certain threshold considering the noise level, the m + 1-th and m + 2-th spectrum amplitudes are replaced by the average value of the two, and thus (B) and (C in FIG. ), The spectral noise can be removed.

The voiced strength determination unit 109 includes the frame energy calculation unit 103 and the band energy calculation unit 10.
4. The frame energy (frame power) Ef and the band energy (band power) Eb [k] calculated by the logarithmic conversion unit 105, the band harmonics amplitude calculation unit 106, the band harmonics width calculation unit 107, and the band harmonics number calculation unit 108. , Harmonics amplitude Hpw [n]
[0], Harmonics width Hpw [n] [1], and Harmonics number Hn are used to calculate and output voiced strength V [k] for each band. Here, Hpw [n] [0] is the amplitude of up to n of the harmonic amplitudes (Ah or Bh) in that frequency band, and Hpw [n] [1] represents the corresponding harmonic width.

The voiced strength V [k] may be a binary voiced / unvoiced determination result obtained by thresholding the input parameter, or has a multivalued level by weighted addition of the input parameter determination values. The judgment result may be used. Alternatively, it may be a binary determination result obtained by thresholding the weighted addition result of the determination values of the input parameters. Voiced strength V
When a binary determination result is used as [k], voice synthesis is performed by switching between voiced and unvoiced for each band. In the case of a multi-valued determination result (for example, a value in the range of 0.0 to 1.0), if voiced and unvoiced synthetic voices synthesized for each band are weighted and added to generate a final synthetic voice, Good.

FIG. 10 and FIG. 11 are processing flow charts showing the processing contents of the voiced strength calculation section 404 in FIG. When the voiced strength calculation is started, the fundamental frequency ωo and the frequency spectrum amplitude | A [m] | are received in step 1401 and set in a data area in 1402. Here we use the fundamental frequency ωo, which is used to determine the number of bands and the frequency range of the bands,
It is not used directly to judge voiced strength. In step 1403, the number of bands K is determined. However, if each band is set to include h harmonics, the number of bands K
Is

[Equation 20] Calculated by Here, ceil (x) represents the smallest integer greater than or equal to x. For example, the number of bands K is calculated by designing about h = 3. Once h and ω o are determined, the frequency region [ak, bk] of the FFT spectrum that falls within each band is calculated for each band number k = 1, 2, ..., K by the above equation (15).

In step 1404, the frame power E
f and band power Eb [k] (k = 1,2, ..., K) are calculated from the equations (13) and (14).

[Equation 21]

[Equation 22] Next, in step 1405, the logarithmic amplitude LA [m] obtained by taking the logarithm of the spectral amplitude | A [m] | and converting it into decibels is calculated.

[Equation 23] Next, in 1406, spectrum noise removal is performed. This spectrum noise removal processing flow (steps 1421 to 14)
28) will be described later.

Next, the voiced strength V [k] is determined. First, in step 1407, a frame in which the power (frame power) Ef of the entire frame is smaller than a predetermined threshold value Th0 is a place where the voice power is small and is considered to be a noise region. Therefore, in step 1416, all bands are set to be unvoiced. Exit without entering a loop. On the other hand, for a frame whose frame power Ef is larger than the threshold Th0, the band loop of steps 1408 to 1415 is entered. In this band loop, first, in step 1409, the power Eb [k] of the frequency band is evaluated, and if it is less than or equal to a predetermined threshold Th1, it is determined that the band has little energy, and unvoiced V [k] = 0. It is set (step 1414). Threshold Th1
If it is larger, in step 1410, the band harmonics amplitude Hpw [n] [0], the harmonics width Hpw [n] [1], and the harmonics number Hn are calculated. In the flowchart, the harmonic amplitude Hpw [n] [0] and the harmonic width H
pw [n] [1] is collectively expressed as Hpw [n] [2]. The processing flow for calculating the harmonic amplitude, the harmonic width, and the number of harmonics in step 1410 (step 14
30-1450) will be described later.

Next, in step 1411, the number of harmonics Hn is evaluated, and if there is a difference from the set number of harmonics in the band h, that is, outside the range (threshold value Th20 or less, threshold value th21 or more), then voiceless V [k] = 0. The determination is made (step 1414). For example, when the number of harmonics h per band is set to 3, 2 or less and 4 or more are determined to be unvoiced sound. Next, in step 1412, the harmonics amplitude Hpw [n] [0] and the harmonics width Hpw [n] [1] are evaluated, and predetermined threshold values Th are obtained.
3, if it is smaller than Th6, the harmonics amplitude is small,
It is determined that the voice has a narrow band width (step 141).
4). The threshold of the harmonics width is adaptively set by the window function set by the adaptive window processing unit 101. For example, in the case of the adaptive Hamming window of the formula (8), it is appropriate to consider the harmonics width in association with the spectrum width of the main lobe represented by the positive and negative first valley distances of the adaptive Hamming window spectrum distribution. Spectral width Mw of the main lobe of the Hamming window by window length L _w and FFT the number of stages M (2
Since it is calculated by the equation (1), Th6 sets a practical threshold value in relation to this value.

[Equation 24] Similarly, Th3 is related to the first attenuation amount of the Hamming window, and if the fundamental frequency and the window length of the adaptive window processing unit satisfy the condition of the above equation (6), the first Hamming window Set a practical value based on the amount of attenuation in the valley. The band not determined as unvoiced sound is set as a voiced band (V [k] = 1) in step 1413. When the above operation is calculated up to the maximum K bands for each band and the setting of each voiced strength V [k] is completed, the processing of the voiced strength calculation unit 404 is finished in step 1417.

In this way, the threshold value is determined for the frame power Ef (1407), and the threshold value is determined for the band power Eb [k] of each band (1409).
A threshold judgment is made for the harmonic number Hn (1411),
Further, threshold determination (1412) is performed on the harmonics amplitude Hpw [n] [0] and the harmonics width Hpw [n] [1], and a binary (0 or 1) voiced strength V is obtained from these determination results.
[k] can be determined. Note that, as described above, the voiced strength V [k] is not limited to such binary information, and a predetermined weight is given to each of the results of the threshold determinations, and these are added. Thus, a multivalued (for example, a range of 0.0 to 1.0) voiced strength may be calculated. Alternatively, the weighted addition result may be determined using a predetermined threshold value and set to a binary value.

Next, the processing contents of the spectral noise elimination subroutines 1421 to 1428 in step 1406 will be described. For the logarithmic value LA [*] of the spectral amplitude received in step 1421, step 142
The noise removal loop of 2-1427 is entered. In this noise elimination loop, it is checked whether or not there are small maximum points in the frequency spectrum amplitudes of four consecutive points.
If there is a small maximum point, the average with the spectrum amplitude having the closest amplitude value to the maximum point is averaged, and the spectrum amplitudes of both are replaced by the average value to eliminate the small spectrum maximum point.

First, at step 1423, the differences d1, d2, d3 of four consecutive points are calculated, and their codes s1, s are calculated.
2. Calculate s3. Next, in step 1424, it is determined whether s1 and s3 have the same sign and are different from s2. When the result is true, the maximum point is either of the two middle points. As shown in FIG. 9, the amplitudes of the maximum points are s1 and s2.
Are both positive and both are negative, they are represented by the same absolute value of d2. In step 1425, | d2 |
If it is smaller than 4, in step 1426, LA [m + 1] and LA [m + 2] are replaced with their average values to smooth out a small maximum value. The above smoothing process is repeated in the band until the spectra of the last four points are obtained, and the maximum points due to spectrum noise are removed. From FIG. 9, it is understood that if the maximum point is removed, the minimum point immediately before or immediately after the maximum point can be obtained at the same time.

Next, the processing contents of the calculation subroutines 1430 to 1450 for the number of harmonics Hn, the amplitude of harmonics Hpw [n] [0], and the width of harmonics Hpw [n] [1] in step 1410 will be described with reference to FIG. . First, step 1
At 431, the logarithmic spectrum amplitude LA [m], the fundamental frequency ωo,
Band number k (k = 1,2, ..., K), band spectrum range
The processing is started by inputting [ak, bk]. Step 14
32, a maximum value number counter Npk for counting the number of maximum values;
A minimum value counter Nbtm that counts the number of minimum values, a maximum value memory Apk [*] that stores the amplitude of the maximum value, a minimum value memory Abtm [*] that stores the amplitude of the minimum value, and a harmonics that stores the amplitude of harmonics. Amplitude memory Hpw [*] [0], harmonic width memory Hpw that stores the bandwidth of harmonics
[*] [1], The harmonics number counter Hn that counts the number of harmonics is initialized to 0. Further, the starting point mb1 and the ending point mb2 of the harmonics width are set to the spectrum starting point ak of the band.

Next, in step 1433, the peak / bottom calculation loop (steps 1433 to 1448) is entered, and in step 1434, the logarithmic spectrum amplitude LA [m] is LA [m-.
If it is larger than 1] and LA [m + 1], it is determined that LA [m] is the maximum value and the process proceeds to step 1435. Step 1435
Then, it is detected whether or not the detected maximum value is the first detected value in the band. If detected, the maximum value number counter Npk and the minimum value number counter Nbtm are set to 1 in step 1436. , The maximum value LA [m] is stored in the maximum value memory Apk [1], and the initial value LA [ak] is stored in the minimum value memory Abtm.
Record in [1]. If it is not the first detected,
In step 1437, the maximum value number counter Npk is incremented, and the maximum value LA [m] is recorded in the maximum value memory Apk [Npk].

On the other hand, when it is determined in step 1434 that the peak is not detected, it is subsequently determined in step 1438 whether it is the minimum value. This determination is performed by the same method as the maximum value determination in step 1434, and when it is determined as the minimum value as a result, the minimum value number counter Nbtm is incremented in step 1439 and the minimum value LA [m] is minimized. The value is stored in the value memory Abtm [Nbtm]. In addition, mb1 is replaced by mb2 to calculate the harmonics width.
And the current spectrum frequency m is set to mb2. If the maximum value and the minimum value are both judged as No,
In step 1441, it is determined whether it is the end of the bottom / peak detection loop, and if it is the last loop, step 144
In step 2, it is determined whether the maximum value number counter value Npk and the minimum value number counter value Nbtm are the same. If they are the same, at step 1440 the minimum value number counter Nbtm is incremented and LA is stored in the minimum value memory Abtm [Nbtm].
Record [bk] and set mb1 to mb to calculate the harmonic width.
Update to 2 and set the final spectral frequency bk of the current band to mb2. This procedure detects all local maxima and records the local minima before and after it.

Next, in step 1443, it is judged whether or not there is a local maximum value before the local minimum value is detected, and if there is, the local maximum value is used as a new harmonics to calculate its amplitude Ha in step 1444. To do. Step 1444
Then, the average value of the amplitude difference between the maximum value and the minimum values before and after the maximum value is defined as the harmonic amplitude Ha. However, if we consider the symmetry of the harmonics amplitude shape as important,
As shown in the above equation (18), Ha may be calculated with the minimum value. Next, in step 1450, the harmonics width Hw
Is calculated, and in step 1445, Ha is compared with a predetermined threshold value Th5.
Is updated (step 1446), and the harmonics width Hpw [n] [1] is recorded in the upper n harmonics amplitudes Hpw [n] [0] (step 1447). MaxN of step 1447
(Hpw [n], Ha, Hw) replaces the minimum value of the first array element that shows the harmonics amplitude when Ha is larger than the minimum value of the array element of Hpw [n] [0], and at the same time sets the harmonics width. The function which replaces the 2nd element of the shown sequence number with Hw is shown. When all the peak / bottom calculation loops have been completed, in step 1449, the number Hn of harmonics in the band, the amplitudes of the top n harmonics and the width Hpw [n] [2] are returned. The processing contents of the voiced strength calculation unit 404 have been described above with reference to the detailed flow chart.

Next, the unvoiced speech synthesizer 509 in the speech decoding apparatus to which the speech decoding method of the present invention shown in FIG. 2 is applied will be described in detail. As described above, the symmetric random sequence generation unit 201 generates a central symmetric random sequence, and the antisymmetric random sequence generation unit 202 generates a central antisymmetric random sequence. Here, the centrally symmetric random sequence is symmetrical in both amplitude polarities with respect to a certain point (center) in the sequence (that is, when folded back at the center, the sequences on the left and right sides of the center are completely matched. State)
The central anti-symmetric random sequence is a random sequence whose amplitude is bilaterally symmetrical from the center but whose polarity is reversed. Actually, a random sequence having a length of 1/2 of the number of stages of inverse Fourier transform processing executed in the inverse frequency transform unit 204 (referred to as the number of inverse FFT stages) is generated, and this is copied in the reverse direction of the generation order. By doing so, it is possible to generate the centrally symmetric random sequence, and generate a random sequence of ½ of the inverse FFT stage number, and invert the polarity in the reverse direction of the generation order to copy the random sequence. , The central antisymmetric random sequence can be generated.

The two random sequences generated by the symmetric random sequence generating unit 201 and the antisymmetric random sequence generating unit 202 in this way are the random sequence extracting unit 20.
3 where the two random sequences are regarded as the real part and the imaginary part of the frequency spectrum sequence, and the sequence of the section corresponding to the frequency band designated as unvoiced by the voiced / unvoiced information is extracted. At the same time, the amplitude is adjusted so that the extracted spectrum power becomes the same as the power of the unvoiced harmonics band corresponding to the spectrum envelope information B | (ω) |. The amplitude-adjusted unvoiced harmonics band spectrum is inverse-Fourier-transformed by the inverse frequency converter 204 to be converted into a signal in the time domain, and the same number of time axis data sequences as the number of stages of the inverse FFT corresponding to the unvoiced frame signal are obtained. To be

The thus obtained, for example, 256 pieces of data (when the number of inverse FFT stages is 256) is input to the frame interpolating unit 205, and the number of data corresponding to the update period of the audio segment (for example, 20 msec). If it is a cycle, 160 pieces of data) are interpolated and synthesized. This is for linearly interpolating the time axis data obtained from the previous segment and the time axis data of this segment under the condition that the sum of the interpolation weights becomes 1. The unvoiced voice synthesized in this way is supplied to the adder 507 and added to the voiced voice from the voiced voice synthesis unit 508 described above.

FIG. 12 is a diagram showing a processing flow of the above-mentioned unvoiced voice synthesis. First, in step 1602, the spectrum decoding information of harmonics | B from the parameter decoding unit.
(ω) |, voice fundamental frequency ωo, band voiced unvoiced information V [k]
Is received and the number of bands Kmax is reproduced by the equation (20). Here, the number of harmonics h included in each band is predetermined by the system. The frame size Fsize is a preset audio segment update interval, fs = 800
Fsize when 0Hz and segment update cycle is 10msec
= 80. Step 1603 is 2 for IFFT stage number M.
When 56 is used, the real part and the imaginary part of the FFT spectrum number are each initialized to 256 elements. Step 160
Reference numeral 4 is an initialization for generating a random FFT spectrum, which is necessary only when the system is started up and not necessary when reproducing continuous voice.

Steps 1605 to 1614 are loops that are processed by the number of bands of the processing frame. A loop for reproducing and adding the spectrum of unvoiced voice in the frequency band of the unvoiced band for each harmonics band to reproduce the entire unvoiced sound spectrum of the frame. Is. In step 1606, a random sequence whose number of elements is half the number of IFFT stages is sequentially generated. For example, in the IMBE method, the random sequence is generated by Expression (22), and this method may be generated by the same method. However, here, a random sequence u [n] of two series for the real part and the imaginary part is generated, and u [n] is a value obtained by subtracting 53125/2 in order to remove the DC component.

[Equation 25]

Steps 1607 to 1613 are a harmonics loop, and are processed by the number of harmonics contained in each band. First, in step 1608, the spectral range of the 1st order harmonics in each band [a
l, bl] is calculated by the equation (23), and step 1609 is performed.
Then, that range is extracted from the random sequence u [n] and u [al, bl] is extracted.

[Equation 26] Here, M is the number of stages of inverse FFT (IFFT).

Next, the spectrum is normalized by the equation (24) so that the power of the extracted spectrum extracted in step 1610 becomes 1. Where U (m) is a vector representation of random sequences u_real [m] and u_imag [m] of real and imaginary parts, and U1 (m) is a normalized extracted spectrum u_Real [m] and u_Imag. This is a vector representation of [m].

[Equation 27] Step 1611 is the spectrum envelope information of harmonics.
By | B (ω) |, the amplitude of the extracted spectrum is adjusted by the equation (25) so that the energy in the harmonics band becomes the same as the energy in the band of the original voice.

[Equation 28] Here, M in the last term is the M stage IF in step 1616.
It is a coefficient necessary for the output of the FT to match the real-time signal level. Next, in step 1612, the level-adjusted extracted spectrum is converted into the corresponding FFT spectrum buffer S_.
Set to real [M], S_imag [M].

After the above processing is executed for each band and each harmonics in each band, the process proceeds to step 1615 to set a negative frequency FFT spectrum portion satisfying the relation of the above equation (10), and at step 1616. M-stage IF
All spectral energy is collected in the real part of the time axis signal obtained by FT, and no signal appears in the imaginary part. In step 1617, the unvoiced decoded speech of the frame sample number (Fsize) is obtained from the signals of the sample number M obtained from the current frame and the previous frame by interframe interpolation of the interpolation function ws (n) shown in Expression (26). Then, in step 1618, the unvoiced decoded speech is reported and added to the speech of the voiced speech synthesis section 508 separately synthesized by the addition section 507 of FIG. 2 to obtain the final synthesized speech. FIG. 13 is a diagram showing an example of the frame interpolation function ws (n). Here, L1 is a constant level range of the interpolation function, L2 is a maximum interpolation range, and L1 to
Between L2 is a linear interpolation range.

[Equation 29]

In the above description, the case where the speech coding parameter acquisition method of the present invention is applied to the speech coding parameter extraction unit of the speech coding transmission apparatus that employs the IMBE method as the speech coding method will be described as an example. However,
A speech coding parameter extraction method and apparatus of the present invention are
The MELP (Mixed Excitati) is not limited to this.
The method can be applied in exactly the same manner when dividing a frequency spectrum of one frame into a plurality of frequency bands and determining voiced / unvoiced for each frequency band, such as on linear prediction. Further, the case where the present invention is applied to the unvoiced voice decoding unit of the voice encoding / transmission device that employs the IMBE system as the voice decoding system has been described as an example, but the unvoiced voice decoding method and device of the present invention are not limited to this. Alternatively, the present invention can be similarly applied to the case of dividing the frequency spectrum of one frame into a plurality of frequency bands and determining voiced / unvoiced for each frequency band, such as the MELP method.

[0065]

As described above, according to the speech coding parameter acquisition method and apparatus of the present invention, the adaptive variable length window is adapted so that the speech harmonics spectrums are separated from each other by the speech fundamental frequency. The frequency spectrum is obtained from the processed speech segment to increase the reliability of the detected harmonics amplitude, harmonics width and number of harmonics, and the power of the speech segment and the power of each frequency band obtained by dividing the speech segment into multiple frequency bands. As a result, voiced strength or voiced / unvoiced information is acquired. Therefore, it is possible to perform voiced strength judgment that is less affected by harmonics harmonic noise to low harmonics levels regardless of changes in the fundamental frequency of the voice. Become. Therefore, it is possible to provide a method for acquiring a speech coding parameter that is highly error resistant to spectrum noise.
Further, according to the speech decoding method and apparatus of the present invention, in decoding unvoiced speech, it is possible to directly generate a random frequency spectrum from random noise by FFT without generating a random frequency spectrum. Of the FFT and IFFT calculations, the IFFT
An unvoiced voice can be synthesized only by calculation, and a voice decoding method with a smaller calculation load than the conventional method can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a speech coding parameter extraction unit to which a speech parameter acquisition method of the present invention is applied.

FIG. 2 is a block diagram showing a configuration of a speech synthesis unit to which a speech decoding method of the present invention is applied.

FIG. 3 is a diagram showing an example of a log spectrum amplitude (voiced part) of a voice segment.

FIG. 4 is a diagram showing an example of a log spectrum amplitude (unvoiced part) of a voice segment.

FIG. 5 is a diagram for explaining a spectrum shape of harmonics.

FIG. 6 is a diagram showing a setting example of a first frequency analysis window length.

FIG. 7 is a diagram showing an example of the shape of a first frequency analysis window.

FIG. 8 is a diagram for explaining harmonics amplitude.

FIG. 9 is a diagram for explaining spectrum noise removal.

FIG. 10 is a flowchart showing a flow of voiced strength calculation processing.

FIG. 11 is a flowchart showing a flow of harmonics calculation processing.

FIG. 12 is a flowchart showing the flow of unvoiced voice synthesis processing.

FIG. 13 is a diagram for explaining inter-frame interpolation of unvoiced voice.

[Fig. 14] Fig. 14 is a diagram for describing the configuration of a voice coding and transmission device.

FIG. 15 is a block diagram of a conventional speech coding parameter extraction unit.

FIG. 16 is a block diagram of a conventional speech synthesizer.

FIG. 17 is a diagram for explaining a relationship between a normalized spectrum error and a pitch frequency error.

[Explanation of symbols]

101 Adaptive window processing unit 102 first spectrum calculation unit 103 Frame energy calculator 104 Band energy calculator 105 logarithmic converter 106 band harmonics amplitude calculator 107 Band harmonics width calculator 108 Band harmonics number calculator 109 Voiced strength determination unit 110 Fixed window processing unit 111 Second spectrum calculation unit 112 Spectrum Envelope Calculator 201 Symmetric random sequence generator 202 Antisymmetric random sequence generator 203 random sequence extraction unit 204 inverse frequency converter 205 frame interpolator

─────────────────────────────────────────────────── --- Continuation of the front page (56) Reference JP-A-7-295593 (JP, A) JP-A-8-272398 (JP, A) JP-A-10-124092 (JP, A) JP-A-7- 261796 (JP, A) Special table Sho 62-502572 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 19/02

Claims (6)

(57) [Claims] 1. A voice coding parameter acquisition method for acquiring a voice coding parameter from a voice segment obtained by extracting a digitized voice signal at a predetermined repetition period at a predetermined segment length, wherein Acquiring a voice fundamental frequency from a segment, acquiring a first frequency spectrum from a variable length segment obtained by extracting the voice signal by a variable length adaptive window determined by the voice fundamental frequency, fixing the voice signal to a fixed length Acquiring a second frequency spectrum from the fixed-length segment extracted by the window, dividing the first frequency spectrum into a plurality of frequency bands, frequency spectrum power of the first frequency spectrum, each frequency band Frequency spectrum power of each frequency band Determining the voiced strength for each frequency band according to the number of harmonics included, the harmonics amplitude of each harmonics, and the harmonics bandwidth; and centering a frequency that is an integral multiple of the voice fundamental frequency from the second frequency spectrum. A method of acquiring a speech coding parameter, comprising the step of calculating the spectrum power of each harmonics band divided so that the frequency bandwidth becomes a speech fundamental frequency. 2. The length of the variable-length adaptive window is determined by the relationship between the bandwidth of the frequency spectrum distribution of the variable-length adaptive window and the voice fundamental frequency. Method for obtaining the speech coding parameters of. 3. The speech coding parameter according to claim 1, wherein the variable-length adaptive window is a Hamming window having a length four times or more of a cycle corresponding to the speech fundamental frequency. Acquisition method. 4. A voice fundamental frequency of a voice segment obtained by extracting a digitized voice signal at a certain repetitive cycle, and a frequency spectrum of the voice segment whose frequency bandwidth is centered at an integral multiple of the voice fundamental frequency. Speech coding that consists of spectrum power of each harmonics band divided to become a voice fundamental frequency and discrimination information that discriminates whether each frequency band obtained by dividing the frequency spectrum of the voice segment into a plurality of frequency bands is voiced or unvoiced A voice decoding method for synthesizing voice according to a parameter, wherein in the frequency band in which the discrimination information indicates a voice, its center frequency has a frequency that is an integral multiple of the voice fundamental frequency, and the spectrum of the corresponding harmonic band. Generate a sine wave group with an amplitude that is equivalent to the power, and In the frequency band different information indicates unvoiced regards centrosymmetric random sequence and the center antisymmetric random sequence and the real and imaginary parts of the frequency spectrum sequence of the noise signal, the 2
An interval corresponding to the frequency band is extracted from the two random sequences, and the amplitude is adjusted so as to be the same as the spectral power of the corresponding harmonic band, and then the real part thereof is obtained by inverse Fourier transform to obtain an unvoiced frame signal. A voice decoding method for generating a voiceless voice by linearly interpolating between the voiceless frame signal of the previous segment and the voiceless frame signal obtained this time, and then adding the generated sine wave group to obtain a synthesized voice. 5. A voice coding parameter acquisition device for acquiring a voice coding parameter from a voice segment obtained by extracting a digitized voice signal at a predetermined repetition period at a predetermined segment length, Means for obtaining a voice fundamental frequency from a segment; means for obtaining a first frequency spectrum by a variable length segment obtained by extracting the voice signal by a variable length adaptive window determined by the voice fundamental frequency; Means for obtaining a second frequency spectrum by a fixed length segment extracted by the window of, a means for dividing the first frequency spectrum into a plurality of frequency bands, a frequency spectrum power from the first frequency spectrum, each of the frequency bands Frequency spectrum power of each frequency band included Monikusu number, means for determining a voiced strength for each of the respective frequency bands by harmonics amplitude and harmonics bandwidth of each harmonics,
And a means for calculating the spectrum power of each harmonic band divided from the second frequency spectrum so that the frequency bandwidth becomes the voice fundamental frequency with a frequency that is an integral multiple of the voice fundamental frequency as the center. A device for acquiring a voice coding parameter. 6. A voice fundamental frequency of a voice segment obtained by extracting a digitized voice signal at a certain repetition period, and a frequency spectrum of the voice segment whose frequency bandwidth is centered at an integral multiple of the voice fundamental frequency. Speech coding that consists of spectrum power of each harmonics band divided to become a voice fundamental frequency and discrimination information that discriminates whether each frequency band obtained by dividing the frequency spectrum of the voice segment into a plurality of frequency bands is voiced or unvoiced A voice decoding device for synthesizing voice by a parameter, wherein, in the frequency band in which the discrimination information indicates voiced, the center frequency thereof has a frequency that is an integral multiple of the voice fundamental frequency, and the spectrum of the corresponding harmonic band. A means to generate a sine wave group with an amplitude equal to the power, center Means for generating a noise signal of a so-called random sequence and a central antisymmetric random sequence, means for extracting a section corresponding to the frequency band in which the discrimination information indicates unvoiced from the two random sequences, a noise signal of the extracted random sequence A means for adjusting the amplitude so that the spectrum power becomes the same as the spectrum power in the harmonics band corresponding to the frequency band in which the discrimination information indicates unvoiced, inverse Fourier transforming the amplitude-adjusted random sequence noise signal, Means for generating an unvoiced frame signal, means for generating unvoiced speech by linearly interpolating the unvoiced frame signal of the preceding segment and the unvoiced frame signal of this time, and the generated sine wave group and the generated unvoiced voice A voice decoding device comprising means for adding voices.