JPH0756599A - Wide band voice signal reconstruction method - Google Patents

Abstract

PURPOSE:To reconstruct wide band voice signals (for example 8kHz/band) from narrow band voice signals (for example 300Hz to 3.4kHz) with a good quality. CONSTITUTION:Input narrow band voice signals are LPC analyzed and their spectrum information parameters are obtained (401). These parameters are vector quantized using a narrow band voice signals code book 208 (402). Original voice waveforms corresponding to the code vectors are stored in a representative waveform piece code book 305 equivalent to one pitch for a voiced sound and equivalent to one frame for a voiceless sound for every code number of the code book 208. Using this code book 305 and output code vector numbers of the step 402, voiced sound are waveform synthesized by superimposing pitch synchronization and voiceless sounds are waveform synthesized by randomly using waveforms equivalent to one frame and wide band voice signals are obtained (404). Wide band components other than 300Hz to 3.4kHz are taken out from the wide band voice signals (405), synthesized with the amplified and sampled narrow band signals and wide band voice signals are reconstructed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for generating a wideband audio signal from a narrowband audio signal, and more specifically, a narrowband audio signal currently output by telephone voice, AM radio, etc. The present invention relates to a method capable of enhancing the quality of a wideband audio signal such as that output from a set or FM radio.

[0002]

2. Description of the Related Art Telephone voice will be described as an example of a narrow band voice signal. The spectrum band of the signal that can be transmitted by the existing telephone system is about 300 Hz to 3.4 kHz. The purpose of the conventional voice coding technique is to maintain the voice quality in this telephone band and to minimize the amount of transmission parameters. That is, although it is possible to reproduce the input voice by the conventional voice encoding technique, it is impossible to obtain the voice that exceeds the quality of the input voice. On the other hand, the quality of sounds used in daily life has been improved due to the recent development of audio technology and the development of digital processing, and there is a situation in which the sound quality of voice in the current telephone band cannot be satisfied. As a method to solve this demand, it is possible to discard the existing telephone system and reconstruct a telephone system capable of transmitting a wideband signal, but it is not only a heavy burden on the economy but also the reconstruction. Even if it does, it will take a considerable amount of time.

[0003]

SUMMARY OF THE INVENTION A main object of the present invention is to enable a narrow band voice signal transmitted by effectively utilizing an existing telephone system to be output as a wide band voice signal, for example, a wide band. A wideband voice signal that enables a wideband voice signal to be used regardless of the combination of both telephone systems, even in the situation where a telephone system capable of transmitting the above signal and an existing narrowband telephone system coexist. It is to provide a restoration method.

[0004]

According to the invention of claim 1, the narrow band speech signal inputted in the first step is subjected to spectrum analysis, and the analysis result is prepared in advance in the second step. Vector quantization is performed using the codebook of the band speech signal, and the quantized value is restored in the third step using the codebook of the representative waveform segment prepared in advance to obtain the wideband speech signal.

According to the invention of claim 2, in the invention of claim 1, in the fourth step, the input narrowband speech signal is upsampled to calculate a sampling value, and in the fifth step, A wide-band portion outside the band of the input narrow-band speech signal is extracted from the wide-band speech signal obtained in the third step, and the extracted wide-band portion is added to the sampling value in the sixth step to obtain a wide-band speech signal.

[0006] According to the invention of claim 3, claim 1 or 2
The codebook of the representative waveform element used in the invention is created as follows. That is, the learning wideband speech signal is spectrum-analyzed, the spectrum analysis result is compared with the codebook of the wideband speech signal prepared in advance, and the learning broadband corresponding to the parameter and the spectrum analysis result closest to the parameter of each code is compared. The voice signal is extracted for one pitch for voiced sound and for one frame for unvoiced sound and used as a representative waveform segment of the code.

According to the fourth aspect of the invention, the input narrow band speech signal is spectrally analyzed in the first step, and the analysis result is used in the second step by using a code book of the narrow band speech signal prepared in advance. Vector quantization is performed, and the quantized value is restored in a third step using a codebook of a wideband speech signal prepared in advance to obtain a wideband speech signal. Component is extracted in the fourth step to restore the low frequency signal,
In the fifth step, the quantized value obtained in the step is used to restore the high-frequency signal in the fifth step using the codebook of the representative waveform segment of the audio signal in the higher-frequency band than the narrow-band audio signal prepared in advance. The banded voice signal is up-sampled in the sixth step to calculate a sampling value, and the sampled value, the restored low band signal and the restored high band signal are added in the seventh step to obtain a wide band voice signal. obtain.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION The method of the present invention requires a codebook of a narrow band speech signal and a codebook of a representative waveform segment,
The codebook of the wideband speech signal is used when creating the codebook of the representative waveform segment. Therefore, first, a method of creating these codebooks will be described.

First, a procedure for creating a codebook of wideband audio signals will be described with reference to FIG. This creation procedure is known in the related art, and in order to efficiently express the characteristics of the wideband speech signal, clustering is performed using parameters that appropriately express the characteristics of the wideband speech signal to create a codebook. As a parameter that characterizes a speech signal, a speech spectrum envelope by linear predictive analysis (LPC) or an FFT
Speech spectrum envelope by Pepstrum analysis, PSE
A voice analysis / synthesis method, a voice expression method by superposition of sine waves, and the like are conceivable. In this embodiment, a case where a voice spectrum envelope by LPC is used as a characteristic parameter will be described. First, the input wide band, for example, 8 kHz band voice is converted into a digital signal by the A / D converter in step 101. Then, in step 102, LPC analysis is performed, and parameters of spectral information (autocorrelation coefficient, LPC cepstrum coefficient) are obtained. After collecting enough of these parameters, for example about 200 words, step 103
Clustering in. Clustering is LB
The G algorithm is used, and the distance measure used at this time is the Euclidean distance D of the LPC cepstrum as shown in the equation (1).

D = Σ [C (i) −C ′ (i)] (1) where Σ is from i = 1 to p, and C and C ′ are different audio signals obtained by LPC analysis. LPC cepstrum coefficient, p is the order of the LPC cepstrum coefficient. In addition,
For the LBG algorithm described above, Linde, Buzo, Gra
y; "An algorithm for Vector Quantization Design" IE
See EE COM-28 (1980-01) in detail.

The wideband speech signal codebook 104 is obtained based on the above equation (1). Next, referring to FIG.
A procedure for creating a narrowband speech signal codebook while taking correspondence with the wideband speech signal codebook will be described. First, in step 201, a signal in the same band as the input narrowband audio signal is created from the learning wideband audio signal. In this embodiment, the wideband voice signal will be described as an 8 kHz band voice signal, and the narrowband voice signal will be described as a telephone band voice signal. Therefore, step 20
1 is realized by passing a wideband audio signal through a high-pass filter that removes frequencies below 300 Hz and a low-pass filter that removes frequencies above 3.4 kHz. On the other hand, the input wideband audio signal is set to LP in step 202.
C analysis is performed, and in step 203, the above-mentioned FIG.
Vector quantization is performed using the codebook 104 of the wideband speech signal obtained according to the codebook creation procedure shown in FIG.

By the way, since the narrow band voice signal is created from the wide band voice signal, the time correspondence between the narrow band voice signal and the wide band voice signal should be one-to-one correspondence with the frame number for which the LPC analysis is performed. You can According to this principle, a narrowband speech signal corresponding to the wideband speech signal vector-quantized in step 203 is obtained, this signal is subjected to LPC analysis in step 205, and the analysis result is analyzed in step 2
At 06, the code vector numbers obtained by the vector quantization in step 203 are classified and stored. That is, matching the time correspondence between the wideband speech signal and the narrowband speech signal and the correspondence between both frames in steps 202 and 205, the vector quantized code vector number of the wideband speech signal having the same frame number, and the narrowband speech signal. The results of the LPC analysis are stored in association with each other. As described above, the processing from step 201 to step 206 is performed on all wideband speech signals prepared for learning, for example, 200 words. In step 207, the LPC analysis result stored in step 206 through all the above processing is averaged for each cluster (same code vector number), and the narrow band speech signal having the average value as a code vector is processed. A codebook 208 is created.

Next, a procedure for creating a codebook of representative waveform segments will be described with reference to FIG. The learning wideband speech signal used for creating the codebook of the wideband speech signal of FIG. 1 or the learning wideband speech signal of different substantially the same frequency band is input and converted into a digital signal by the A / D converter in step 301. Then, in step 302, LPC analysis is performed on the digital signal to obtain parameters of spectrum information (autocorrelation coefficient, LPC cestrum coefficient). This parameter is the same as the parameter used when the codebook of FIG. 1 was created. Therefore, steps 301 and 302 of FIG. 3 are omitted, and step 102 of FIG.
You may use the parameter obtained by. After collecting enough parameters, for example, about 200 words, in step 303, referring to the codebook 104 of the wideband speech signal created in FIG. 1, the frame closest to each centroid is selected, that is, , For each code number in the codebook 104, the spectral information parameter closest to the code vector is selected, and the waveform of 1 analysis window length (in LPC analysis) in the original wideband speech signal corresponding to the parameter is If 1
If it is an unvoiced sound corresponding to the pitch, one analysis window length is extracted, and the extracted waveform is used as a representative waveform segment for the code number. In this way, the representative waveform segment codebook 305 accommodating the representative waveform segment for each code number is obtained.

Next, a procedure for restoring a wideband speech signal from a narrowband speech signal according to the invention of claim 1 will be described with reference to FIG. The input narrow band speech signal of, for example, 300 Hz to 3.4 kHz is subjected to LPC analysis in step 401 to obtain spectrum information similar to that used in FIG. 1, and the obtained spectrum information parameter is set in step 401.
Vector quantize at. In this vector quantization, the narrowband speech signal codebook 208 created by the method shown in FIG. 2 is used. Next, in step 404, the representative waveform segment corresponding to the code vector number is extracted using the representative waveform segment 305 obtained in FIG. 3, the voiced part is a pitch-synchronized superposition, and the unvoiced part is the window length shift width. Waveform synthesis is performed by randomly using the waveforms (in the LPC analysis) to obtain a wideband speech signal in the 8 kHz band, for example.

The wideband speech signal obtained by the above processing is
Although it contains a signal that does not exist in the narrow band audio signal, it has the side effect of distorting the signal that was present in the narrow band audio signal. Therefore, in the invention of claim 2, the following processing is performed. That is, in step 405, the wideband speech signal obtained in step 404 is passed through a lowpass filter for extracting frequencies below 300 Hz and a highpass filter for extracting frequencies above 3.4 kHz, and a wideband portion outside the band of the input narrowband speech signal is obtained. Take out. On the other hand, the input narrow band speech signal is up-sampled to the 8 kHz band in step 406, and the sampling value and step 40
In step 407, the output wideband portion of 5 is added to obtain a restored wideband speech signal.

In the above description, the restoration outside the input signal band by the representative waveform segment codebook is performed only on the high frequency side, and the speech segment need not be used on the low frequency side. That is, for example, as shown in FIG. 5, as described with reference to FIG. 4, the input narrowband speech signal is subjected to LPC analysis (401), and the analysis result is vector-quantized using the narrowband speech signal codebook 208 (402). ), The quantized code is restored using wideband speech signal codebook 104, that is, wideband speech signal codebook 1 having the same code as this quantized code.
The code vector of 04 controls the filter coefficient of the LPC speech synthesizer, inputs the excitation signal of the pitch based on the LPC analysis in step 401 to the LPC speech synthesizer, and sets the output level to the level of the LPC analysis in step 401. It is controlled accordingly (501). From the wideband speech signal thus restored, the low-frequency component, for example, 30
A value of 0 Hz or less is taken out by a low pass filter (502). At this time, the detected power by the LPC analysis in step 401 is the power of the input narrow band signal of 300 Hz to 3.4 kHz, and this power and the LPC composite signal from step 501, for example, the signal in the 8 kHz band has the same power. To be. Therefore, the power level of the restored wideband speech signal is lowered in the band of the input narrowband signal. The reduced amount is step 50
Adjust the output of the low-pass filter of 2 and raise it so as to correspond to the power level of the input narrowband speech signal (5
03). In this way, a signal in a band lower than that band corresponding to the input narrow band voice signal is restored.

Next, two representative waveform segment codebooks are created in advance in order to restore a signal in a higher band than the band corresponding to the input narrow band speech signal. That is, as shown in FIG. 3, in the same manner as in the case of forming the representative waveform segment codebook 305 from the learned wideband speech signal, the learning wideband speech signal is converted into the wideband speech signal codebook 104 as shown in FIG. Vector quantization is performed using the code vectors of the respective codes, and the waveform of the nearest learning wideband speech signal is extracted for one pitch for voiced speech and for one analysis window length in LPC analysis for unvoiced speech (30
3). Each extracted representative waveform segment is, for example, 300 Hz
Pass through a filter with a pass band of ~ 3.4kHz (60
1) Create a narrow-band representative waveform segment codebook 602, and pass each representative waveform segment extracted in step 303 through a high-pass filter that passes, for example, 3.4 kHz or more (603). Representative waveform element codebook 60
Make 4.

Both of these representative waveform element codebooks 60
A method of recovering a signal in a band higher than the band from the input narrow band audio signal using the H.2 and 604 will be described with reference to FIG. In step 402 of FIG. 4, the input narrow band signal is
Vector quantization was performed using the narrowband speech signal codebook 208. Using the quantized code number, a representative waveform segment was selected from the narrowband representative waveform segment codebook 602 (701) and the quantization was performed. Using the code number, a representative waveform segment is selected from the wideband representative waveform segment codebook 604. It is determined whether each of the selected narrowband and wideband representative waveform segments is unvoiced or voiced (7).
03, 704). When it is determined that the voice is unvoiced, the starting point is randomly selected and taken out for each window shift width from each of the selected narrow band and wide band representative waveform segments (70).
5, 706). If it is determined to be voiced sound in steps 703 and 704, the selected narrow-band and wide-band representative waveform segments are extracted in synchronization with the pitch obtained as a result of the LPC analysis (step 401 in FIG. 4). Overlay (70
7, 708). The ratio of the power of each representative waveform segment sequence extracted in steps 705 and 707 and the power obtained as a result of the LPC analysis (step 401 in FIG. 4) is obtained (709 and 710). These calculated ratios are obtained in steps 706 and 708, and are multiplied by the level of the representative waveform segment sequence, and as the power corresponding to the power of the input narrowband speech signal (711, 712), the input narrowing is performed. A restored signal on the high band side of the band corresponding to the band voice signal is obtained (713).

This high frequency side restored signal, the low frequency side restored signal obtained in FIG. 5 and the narrow band speech signal obtained in step 406 in FIG. 4 are added to obtain a wide band speech signal. Step 102 in FIG. 1, Steps 202 and 205 in FIG.
In each spectrum analysis in step 302 in FIG. 3 and step 401 in FIG. 4, the same kind of parameter is obtained by the same analysis method. Narrowband speech signal codebook 20 of FIG.
The wideband speech signal for learning used to create 8 is preferably the wideband speech signal used to create the wideband speech signal codebook 104. In any case, both codebooks may be created while preserving the correspondence relationship between the features of both audio signals. However, although the sound quality is slightly worse than in this case,
A completely different voice signal may be used for each of the codebook 104 for the wideband voice signal and the codebook 208 for the narrowband voice signal, and the codebook 20 for the narrowband voice signal may be used.
8 as shown in FIG. 2, the narrowband speech signal codebook 208 is generated by the usual method shown in FIG. 1 instead of storing the correspondence between the characteristics of the wideband speech signal and the narrowband speech signal. You can make it. Even in this way, the characteristics of the wideband speech signal and the narrowband speech signal are generally highly correlated in terms of the same phoneme, and the same phoneme of the narrowband speech signal is the same in the codebook of the wideband speech signal. It is expected that the sound quality will be considerably improved by using the phonemes. In this case, wideband audio signal codebook 1
It is necessary to associate both code vectors with respect to the same code number of 04 and the codebook 208 of the narrow band voice signal.

A wideband speech signal for creating the wideband speech codebook 104 and a representative waveform segment codebook 305.
Similarly, it may be the same as or different from the wideband audio signal for producing

[0021]

EFFECTS OF THE INVENTION Phonological balance words 18 are used as learning data.
6 words are used, a Hamming window is used as an analysis window, an analysis window length is 21 milliseconds, a window shift width is 3 milliseconds, an LPC analysis order is 14th order, and an FFT point number is 512.
The wide-band speech signal codebook 10 is defined as the LPC cepstrum Euclidean distance is used as the distance measure when the codebook is created.
4 for size 16, narrowband speech signal codebook 206
The following tests were conducted with a size of 256.

Telephone band voice as A and B, 7.3 kHz
Bandwidth is presented randomly. The speech X presented third is set as the following (1) to (4). (1) Telephone band voice (2) 7.3 kHz band voice (3) Voice using the restoration method (present invention) of FIGS. 5 and 7 (4) Low and high frequencies of LPC analysis and synthesis sound of (2) (4) was considered to be the best restored voice when the LPC system was used, and the subject was asked to judge whether the voice of X was close to A or B. A total of 125 voices were presented using headphones. There are 6 test subjects. The following shows the rate at which X was judged to be close to 7.3 kHz band sound. (1) (2) (3) (4) 6.9% 96.9% 86.2% 86.4% The restored voice according to the present invention of (3) and the restored voice by LPC synthesized voice of (4) are 7 86.2% and 86.4%, respectively, were close to the 7.3 kHz band voice, and 86.4% were the restored voice was close to the 7.3 kHz band voice.
%, And it is understood that the band is restored fairly well according to the present invention.

As described above, according to the present invention, it is possible to efficiently restore the characteristics of the voice signal that does not exist in the narrowband voice signal by the correspondence between the characteristics of the narrowband voice signal and the wideband voice. In particular, since the wideband speech signal is used as it is as a speech waveform segment to be restored, high-quality restored speech can be obtained.

[Brief description of drawings]

FIG. 1 is a diagram showing a procedure for creating a wideband speech signal codebook.

FIG. 2 is a diagram showing a procedure for creating a narrowband speech signal codebook.

FIG. 3 is a diagram showing a procedure for creating a representative waveform segment codebook.

FIG. 4 is a diagram showing a procedure of a wideband speech signal restoration method according to the first and second aspects of the present invention.

FIG. 5 is a diagram showing a procedure of a method of restoring a signal in a band lower than that of an input narrowband speech signal.

FIG. 6 is a diagram showing a procedure for creating a narrow-band representative waveform segment codebook and a high-frequency representative waveform segment codebook.

FIG. 7 is a diagram showing a procedure of a method for restoring a signal in a band higher than that of an input narrowband voice signal.

Claims (4)

[Claims] 1. A wideband speech restoration method for generating and outputting a wideband speech signal from an input narrowband speech signal, the first step of spectrally analyzing the input narrowband speech signal, and the first step thereof. The second step of vector-quantizing the result obtained in step 2 using a codebook of a narrow-band speech signal prepared in advance, and the quantized value obtained in the second step are stored in a representative waveform segment prepared in advance. A third step of obtaining a wideband speech signal by restoration using a codebook, and a wideband speech signal restoration method comprising: 2. A fourth step of upsampling the input narrowband speech signal to calculate a sampling value; and a step of calculating the sampling value from the wideband speech signal obtained in the third step, outside the input narrowband speech signal band. And a sixth step of obtaining a wideband speech signal by adding the wideband portion obtained in the fifth step and the sampling value obtained in the fourth step. The wideband audio signal restoration method according to claim 1, wherein 3. The representative waveform segment codebook is obtained by spectrally analyzing a learning wideband speech signal, comparing the spectrum analysis result with a previously prepared wideband speech signal codebook, and regarding each code thereof. A wideband speech signal for learning corresponding to the parameter and the closest spectrum analysis result is set for 1 pitch for voiced sound and 1 for unvoiced sound.
3. The wideband speech signal restoration method according to claim 1, wherein the frame is extracted and used as a representative waveform segment of the code. 4. A wideband speech restoration method for generating and outputting a wideband speech signal from an input narrowband speech signal, the first step of spectrally analyzing the input narrowband speech signal, and the first step thereof. The second step of vector-quantizing the result obtained in step 2 using a codebook of a narrowband speech signal prepared in advance, and the quantized value obtained in the second step are used as a code of a wideband speech signal prepared in advance. A third step of restoring using the book, a fourth step of extracting a low-frequency component from the input narrowband speech signal from the restored wideband speech signal, and a quantized value obtained in the second step. Prepared in advance,
A fifth step of restoring a narrow band speech signal by using a codebook of a representative waveform segment of the high band speech signal to obtain a high band speech signal, and performing upsampling of the input narrow band speech signal The sixth step of calculating the sampling value, the low frequency component extracted in the fourth step, the high frequency audio signal restored in the fifth step, and the sampling value calculated in the sixth step In addition, a seventh step of obtaining a wideband speech signal, and a wideband speech signal restoration method comprising: