JP2005284316A - Method and device for wide-band speech restoration - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a wide-band speech restoring device that restores a wide-band speech signal of high quality by estimating a wide-band sound source which is hardly affected by a difference of a speaker and noise, has a more accurate amplitude, and is stable, from a narrow-band speech or narrow-band speech code. <P>SOLUTION: The speech restoring device is equipped with: a narrow-band sound source decoding means of generating a narrow-band synthesized sound by using the narrow-band speech code, a spectrum decoding means of estimating a wide-band spectrum parameter by using the narrow-band spectrum code separated from the narrow-band speech code; a wide-band sound source decoding means of estimating a wide-band sound source signal by using the narrow-band sound source code separated from the narrow-band speech code; and a synthesizing means of generating a wide-band speech signal from the generated narrow-band synthesized sound, estimated wide-band spectrum parameter, and wide-band sound source signal. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a wideband speech restoration apparatus that restores a wideband speech signal from a narrowband speech signal whose bandwidth is limited or a narrowband speech code obtained by encoding a narrowband speech signal.

  An example of a narrowband audio signal is current telephone audio. In the telephone system, the audio signal is transmitted by being limited to a band of about 300 Hz to 3.4 KHz, and the sound quality is poor and ugly as compared with the case where there is no band limitation. In order to improve the quality, it is conceivable to construct a telephone system capable of transmitting a wide-band audio signal, but much time and cost are required.

  Japanese Patent Laid-Open No. 6-118995 is known as a conventional wideband voice restoration method for restoring a wideband voice signal from narrowband voice limited to a telephone band.

  Japanese Patent Laid-Open No. 6-118995 calculates spectrum parameters by performing LPC analysis on a narrowband speech signal, and vector-quantizes the spectrum parameters using a narrowband codebook. Then, wideband spectral parameters are decoded using the wideband codebook learned in association with the narrowband codebook. An LPC synthesis process is performed using this spectral parameter to obtain a provisional broadband audio signal. A final wideband audio signal is generated by extracting and adding a band component other than the narrowband audio signal from the temporary wideband audio signal to the upsampled narrowband audio signal. Note that, when performing wideband LPC synthesis processing, a wideband sound source signal is required, but a method for generating this sound source signal is not specifically disclosed.

  Reference 1 “Reconstruction of Wideband Speech from Narrowband Speech Using Codebook Mapping” has the same configuration as Japanese Patent Laid-Open No. 6-118995 and is disclosed for generation of a wideband sound source signal, IEICE, IEICE Technical Report. There is SP93-61 (1993-08).

  This document 1 discloses two methods as a broadband sound source generation method.

  The first method uses a pitch and power obtained by analyzing narrowband speech to generate a sound source by a common method among those skilled in the art. That is, an impulse train that repeats with a pitch period is generated for voiced sound, and white noise is generated for unvoiced sound, and its amplitude is determined by power.

  In Document 1, some post-processing is performed to improve sound quality. When restoring a low frequency of 300 Hz or less, the power of the low frequency restoration sound is multiplied by a few times to compensate for the power shortage of the restoration band. When restoring the high frequency range from 3.4 Hz to 7.3 KHz, a cosine function is applied so as to crush the pulse in order to reduce the pulse-like sound generated by using the impulse train as the sound source.

  In the second method, the spectral parameters of the narrowband speech signal are vector-quantized, and a narrowband representative waveform segment and a highband representative waveform segment corresponding to the obtained code are selected. Then, the following processing is performed on the two waveform segments. The voiced voice of the waveform segment is determined, and in the case of voiced sound, the waveform segment is superimposed in synchronization with the pitch obtained by analyzing the narrowband speech signal. In the case of an unvoiced sound, a signal having a required length is cut out from a random position of the waveform segment. The power ratio between the synthesized sound and the narrowband speech synthesized from the narrowband waveform segment using the signal generated by the above process and the narrowband spectrum parameter is calculated. Then, a synthesized sound is generated from the high-frequency waveform segment using the signal generated by the above processing and a broadband spectral parameter, and a high-frequency restored signal is obtained by multiplying this by the power ratio.

  Although the field of use is different, as another method for expanding the band of a sound source signal, reference 2 “A 2.4 Kbps High-Qaulity Speech Coder” IEEE International Conference on Acoustics, Speech, and Signal Processing vol.1, S9.5, pp. 589-592 (1991.5).

Reference 2 relates to a method for efficiently encoding and decoding telephone band speech, and in order to reduce the amount of sound source information during encoding, a long-period prediction analysis of a sound source signal from 0 Hz to 3.4 KHz is performed. And separating into a long-period prediction coefficient and a long-period prediction residual signal. The long period prediction residual signal from 0 Hz to 3.4 KHz is band-limited from 0 Hz to 1 KHz and encoded. Then, after generating a long-period prediction residual signal of the telephone band up to 3.4 KHz from the long-period prediction residual signal whose band is limited when decoding, a long-period synthesis process is performed to restore the sound source signal It is. The long-period prediction residual signal is restored by up-sampling a signal having a component from 0 Hz to 1 KHz to a sampling frequency of 8 KHz, leaving it at an interval of 4 samples, and setting the rest to zero.
Japanese Patent Laid-Open No. 06-118995 Japanese Unexamined Patent Publication No. 63-034600 JP 05-297898 A Japanese Patent Laid-Open No. 06-085607 IEICE Technical Report, SP93-61, 1993

  The above conventional methods have the following problems.

  Japanese Patent Laid-Open No. 6-118995, which is another document, but in Document 1, which discloses a specific practical example, is roughly divided into the following four problems: sound source amplitude estimation, sound source generation method, spectrum parameter estimation method There is a problem related to application to communication systems.

  First, the first sound source amplitude estimation will be described.

  When the first sound source generation method of Document 1 is used, the power value obtained by analyzing the narrowband speech is used as it is or by multiplying the power used for synthesizing the restored sound. Since the gain of the synthesis filter is different between the estimated spectrum parameter and the estimated broadband spectrum parameter, the amplitude of the synthesized sound is different even when the same sound source amplitude is given. Since this difference changes from frame to frame, there is a problem that broadband sound having the correct amplitude cannot be restored by multiplying the sound source amplitude, that is, the power value by a constant.

  In addition, when the second sound source generation method of Document 1 is used, a narrowband synthesized sound is generated and a power ratio with the narrowband speech is calculated and multiplied by the highband synthesized sound. However, there is a problem that requires complicated processing to be executed.

  Next, the second sound source generation method will be described.

  When the first sound source generation method disclosed in Document 1 is used, a wide-band sound source signal is generated using only a small amount of information such as pitch and power, so that the original wide-band sound source that changes in various ways cannot be sufficiently estimated. As a result, although the pulse-like sound is reduced by the cosine function, there is a problem that the sound quality is unnatural because the pulse-like sound cannot be completely suppressed. In addition, since it is impossible to express a voiced sound source having greatly different characteristics for each speaker with one fixed sound source, there is a problem that sound quality deteriorates depending on the speaker.

  When using the second sound source generation method of Document 1, the representative waveform segment corresponding to the sign of the vector quantization result of the spectral parameter is used, but the spectral parameter originally depends on the shape of the vocal tract, and the sound source waveform is Since it depends on how the vocal cords vibrate, there is no strong correspondence between them. The sound source waveform is rather dependent on the speaker. Therefore, there is a problem that an appropriate sound source is not selected.

  As described in Document 1, when this second sound source generation method is used, a waveform segment of an unvoiced sound is selected even though it is a voiced sound. Regardless, there is a case where a waveform segment of voiced sound is selected, and there is a problem in that quality degradation occurs if synthesis is performed as it is. In order to avoid this, the power ratio in that portion is forcibly set to 0, but as a result, the restored high-frequency amplitude becomes partially 0, causing another quality degradation. is there.

  Furthermore, in both sound source generation methods, there is a problem that quality degradation is unavoidable when a voiced / unvoiced determination or pitch extraction error occurs. In particular, when applied to a narrowband audio signal on which noise is superimposed, there is a problem that determination errors and extraction errors increase, resulting in significant degradation.

  Moreover, since there are only two modes of voiced sound and unvoiced sound, a sound source having an intermediate property cannot be expressed sufficiently, and there is a problem that quality degradation occurs at the boundary portion between voiced sound and unvoiced sound.

  Next, a third spectral parameter estimation method will be described.

  In Japanese Patent Application Laid-Open No. 6-118995 and Document 1, vector quantization and inverse quantization using two codebooks are performed. However, a memory for storing the codebooks is necessary, because of quantization processing. The problem is that a large amount of computation is required.

  Further, noise, unvoiced sound, and voiced sound can be easily distinguished by power, and the correspondence between the narrowband spectral parameter and the wideband spectral parameter changes depending on the distinction. However, in any case, since spectral parameters and power are handled independently, information on power is not reflected in the estimation of broadband spectral parameters. For this reason, if the shapes of narrow-band spectra are similar, there is a problem that a similar wide-band spectrum is estimated regardless of the magnitude of power.

  Finally, application to the fourth communication system will be described.

  When applying the method of Japanese Patent Laid-Open No. 6-118995 and Document 1 to a communication system, after decoding a narrowband synthesized sound from the received speech code, the narrowband synthesized sound is reanalyzed to restore the wideband speech signal. However, when spectral parameters and sound source information are transmitted after being separated and encoded, it is considered more efficient to restore the wideband audio signal by directly using the audio code. That is, the method disclosed in Japanese Patent Laid-Open No. 6-118995 and Document 1 has a problem that it is inefficient in that it requires reanalysis. In addition, distortions due to interpolation during synthesis and windowing during analysis are superimposed on parameters obtained by synthesis and reanalysis, and there is also a deterioration in the quality of wideband speech.

  In addition, Japanese Patent Laid-Open No. 6-118995 and Document 1 do not add signal processing that is generally introduced for reducing noise or improving intelligibility of the synthesized sound, so There is a problem that cannot be improved when the sound quality is insufficient.

  In addition, when applied to a communication system, signal processing may be applied to narrowband synthesized sound. In order to superimpose a processed narrowband audio signal and an unprocessed wideband audio signal, There is a problem that continuity of sound quality deteriorates.

  In the method of Reference 2, if a narrow band is considered from 0 Hz to 1 KHz and a wide band is considered from 0 Hz to 3.4 KHz, wide-band sound source signal estimation is performed. As described above, this method inputs a wide-band audio signal. The parameters obtained by analyzing this are encoded and decoded to obtain a wideband synthesized sound, and the wideband speech signal is restored from the narrowband speech signal or the parameters extracted from the narrowband speech signal. It does not disclose how to do this.

  The embodiment described below is made to solve such a problem, and aims to realize a wideband speech restoration apparatus that restores a wideband speech signal having a more correct amplitude from narrowband speech.

  It is another object of the present invention to realize a wideband speech restoration apparatus having a broadband sound source amplitude estimation process of relatively simple processing.

  It is another object of the present invention to realize a wideband speech restoration apparatus that estimates a wideband sound source that is less dependent on a speaker and is good even near a voiced and unvoiced boundary, and restores wideband speech with stable and natural sound quality.

  Another object of the present invention is to realize a wideband speech restoration apparatus that is less affected by voiced / unvoiced judgment errors and pitch extraction errors that tend to occur on narrowband speech signals with superimposed noise.

  It is another object of the present invention to realize a wideband speech restoration apparatus that efficiently restores wideband speech without reanalysis when applied to a communication system.

  Furthermore, when the sound quality of the restored wideband audio signal is insufficient, it can be improved, and when the signal processing is applied to the narrowband synthesized sound, the processed wideband audio signal with good narrowband continuity The purpose is to realize a wideband speech restoration apparatus that can obtain the above.

A wideband speech restoration method according to the present invention comprises a narrowband spectrum decoding step for decoding a narrowband spectrum parameter from a narrowband speech code;
A narrowband excitation decoding step for decoding a narrowband excitation signal from the narrowband speech code;
A broadband excitation decoding step for generating a broadband excitation signal having the same power as the power of the narrowband excitation signal based on white noise;
A spectrum decoding step of outputting a broadband spectral parameter representing a spectral envelope reflecting a shape of a high frequency band of the spectral envelope represented by the narrow band spectral parameter in a high frequency shape;
And a synthesis step for generating a broadband audio signal based on the broadband spectrum parameter and the broadband sound source signal.
A wideband speech restoration apparatus according to the present invention comprises a narrowband spectrum decoding means for decoding a narrowband spectrum parameter from a narrowband speech code,
Narrowband excitation decoding means for decoding a narrowband excitation signal from the narrowband speech code;
Wideband sound source decoding means for generating a wideband sound source signal having the same power as the power of the narrowband sound source signal based on white noise;
Spectrum decoding means for outputting a broadband spectral parameter representing a spectral envelope reflecting a shape of a high frequency band of the spectral envelope represented by the narrow band spectral parameter in a high frequency shape;
And synthesizing means for generating a broadband audio signal based on the broadband spectrum parameter and the broadband sound source signal.

Example 1.
An embodiment of the present invention will be described with reference to the drawings.

  The present embodiment mainly explains the configuration and operation for restoring the generation of the broadband sound source signal in a more correct form.

  FIG. 1 is a configuration diagram of a wideband speech restoration apparatus according to Embodiment 1 of the present invention. In the figure, 1 is an input narrowband speech signal, 2 is analysis means, 3 is spectrum analysis means, 4 is narrowband spectrum parameters, 5 is an inverse filter, 6 is a narrowband sound source signal, 7 is wideband spectrum estimation means, 8 Is a vector quantization means, 9 is a narrowband spectrum codebook, 10 is a spectrum code, 11 is an inverse quantization means, 12 is a broadband spectrum codebook, and 13 is a broadband spectrum parameter. 14 is a wideband sound source estimation means that is an important new component in this embodiment, 15 is a zero padding means as a specific example thereof, 16 is a wideband sound source signal, 17 is a synthesis filter as a synthesis means, 18 is a bandpass filter, 19 is an upsampling means, and 20 is a wideband audio signal.

  FIG. 2 is a signal explanatory diagram for explaining the processing of the zero padding means 15.

  Hereinafter, the operation of the first embodiment of the present invention will be described with reference to FIGS.

  First, for example, a narrowband audio signal 1 sampled at 8 KHz and limited to a telephone band from 300 Hz to 3.4 KHz is input to the analysis unit 2 and the upsampling unit 19. The spectrum analysis means 3 in the analysis means 2 analyzes the narrowband speech signal 1 to calculate a narrowband spectrum parameter 4 and outputs it to the inverse filter 5 and the wideband spectrum estimation means 7 in the analysis means 2. As the narrow band spectrum parameter 4, various parameters such as a linear prediction coefficient, an LSP, a PARCOR coefficient, and a cepstrum can be applied. The inverse filter 5 performs inverse filtering on the narrowband sound signal 1 using the narrowband spectrum parameter 4 and outputs the obtained narrowband sound source signal 6 into the wideband sound source estimation means 14.

  The vector quantization means 8 in the wideband spectrum estimation means 7 vector-quantizes the narrowband spectrum parameter 4 using the narrowband spectrum codebook 9 and the obtained spectrum code 10 is inversely quantized in the wideband spectrum estimation means 7. Output to the conversion means 11. The inverse quantization means 11 inversely quantizes the spectrum code 10 using the broadband spectrum codebook 12 and outputs the obtained broadband spectrum parameter 13 to the synthesis filter 17.

  The processing in the wideband spectrum estimation means 7 is the same as that in Document 1, and detailed description regarding the generation method of the narrowband spectrum codebook 9 and the broadband spectrum codebook 12 and the vector quantization method is omitted.

  The zero padding means 15 in the wideband sound source estimation means 14, which is an important part of the present embodiment, inserts zero M-1 samples between the sample values of the narrowband sound source signal 6, and the number of samples obtained is M times. Is output to the synthesis filter 17 as a broadband sound source signal 16. Here, M is a value obtained by dividing the sampling frequency of the restored wideband audio signal by the sampling frequency of the narrowband audio signal. In this embodiment, a case where M is 2 will be described. FIG. 2A shows an N sample narrowband sound source signal 6. When zero padding processing is performed on this signal by the zero padding means 15, M-1 samples, that is, zeros of one sample are inserted between the samples, and a 2N-sample wide-band sound source shown in FIG. A signal 16 is obtained. When M is zero-padded, a spectrum that is symmetric from 0 Hz to 4 KHz is restored from 4 KHz to 8 KHz, centering on half the sampling frequency of the wideband audio signal, that is, 4 KHz.

  The synthesis filter 17 performs synthesis filter processing on the broadband sound source signal 16 using the broadband spectrum parameter 13 to generate a temporary broadband audio signal. The band filter 18 performs band pass filter processing on the provisional wideband audio signal to extract components other than the band in which the narrowband audio component exists. When the band of the wideband audio signal is from 0 Hz to 7.3 KHz, components from 0 Hz to 300 Hz and 3.4 KHz to 7.3 KHz are extracted.

  The upsampling means 19 upsamples the narrowband audio signal 1 M times. The signal generated by the upsampling has the same narrowband component as the narrowband audio signal 1 with the same sampling frequency as the wideband audio signal 20. Then, the output of the band filter 18 and the output of the upsampling means 19 are added to generate a wideband audio signal 20.

  Originally, the narrowband sound source signal and the wideband sound source signal reflect the characteristics of the sound source signal generated from the same vocal organ, so the strength of the harmonic component of the pitch frequency and the strength of the noise component between the harmonic components There is a correlation in the characteristics of the sound source signal such as. In other words, if the narrowband sound source signal has regular features with strong pitch frequency harmonic components, the wideband sound source signal has the same regular features with strong pitch frequency harmonic components. On the other hand, when the narrow-band sound source signal has a strong noise component, the broadband sound source signal has a strong noise component.

  By configuring the wideband sound source estimation means as in this embodiment, it is possible to generate a 0-8 KHz wideband sound source signal having the same characteristics as the low frequency 0-4 KHz narrowband sound source signal. There is an effect that it is possible to restore wide-band sound with a stable and natural sound quality.

  Also, unlike the conventional example, voiced / unvoiced judgment and pitch extraction are not required, and this structure can naturally represent a sound source with intermediate characteristics, so voiced / unvoiced judgment errors that tend to occur for narrowband speech signals with superimposed noise In addition, it is possible to estimate a good broadband sound source near the voiced and unvoiced boundary without being affected by the error of the pitch extraction and the pitch extraction, and to restore the broadband sound having a stable and natural sound quality.

Example 2
FIG. 3 is a configuration diagram of the sound source estimation unit 14 in the wideband speech restoration apparatus according to the second embodiment of the present invention. The novel parts in the figure are 21 sound source analysis means, 22 narrowband adaptive codebook, 23 distortion minimizing means, 24 narrowband drive sound source signal, 25 narrowband adaptive lag length, and 26 narrowband adaptive gain. 27 broadband drive excitation estimation means, 28 zero padding means, 29 broadband drive excitation signal, 30 broadband adaptive excitation estimation means, 31 broadband adaptive excitation codebook, 32 broadband adaptive excitation signal, 33 broadband adaptive lag Long, 34 wideband adaptive gain. Since the entire configuration is the same as that in FIG. 1, description of the configuration and description of operations other than those in FIG. 3 are omitted.

  According to this configuration, the broadband sound source signal can be restored even better.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  The narrow band sound source signal 6 is input to the sound source analysis means 21 in the wide band sound source estimation means 14. The narrowband adaptive codebook 22 in the sound source analyzing means 21 stores the past narrowband sound source signal 6, and the lag length is an integer value according to the lag length sequentially output by the distortion minimizing means 23 described later. In this case, a signal obtained by repeating the stored past narrowband sound source signal 6 with this lag length is output. When the lag length is a non-integer value, reference 3 “Pitch Predictors with High Temporal Resolution” IEEE International Conference on Acoustics, Speech, and Signal Processing vol.2, S12.6, pp.661-664 (1990.4) A signal is generated and output by the polyphase filter output as described. The length of the output signal is the same as that of the current narrowband sound source signal 6.

  FIG. 4 shows an example of a past narrowband excitation signal 6 stored in the narrowband adaptive codebook 22 and a signal output according to the input lag length.

  In the figure, the horizontal axis indicates that time elapses in the arrow direction. (A1) and (B1) indicate the time length of the sound source signal, (A2) and (B2) indicate the lag length normalized with respect to the output time, such as 20 to 128, ( A3) and (B3) show examples of output sound source signals.

  FIG. 4A shows a case where the length of the output signal is shorter than the lag length. In this case, the source signal (A3) having the output signal time T1 from the beginning of the lag length is continued to the past source signal. Output. When the lag length is shorter than the length of the signal to be output, such as T2, the same sound source signal (B3) is repeated a plurality of times as shown in FIG.

  The distortion minimizing means 23 sequentially outputs a plurality of lag length values to the narrowband adaptive codebook 22, and a signal obtained by multiplying each lag length by a signal output from the narrowband adaptive codebook 22 And the gain of the narrow-band sound source signal 6 are determined so as to minimize the distortion. Then, all the lag lengths that minimize the distortion are selected and output to the wideband adaptive sound source estimation means 30 as the narrowband adaptive lag length 25. Further, the value of the gain at that time is output to the wideband adaptive excitation estimation means 30 as the narrowband adaptive gain 26, and the signal output by the narrowband adaptive codebook 22 is multiplied by the narrowband adaptive gain 26 and the narrowband excitation signal 6. Is output to the wideband drive sound source estimation means 27 as the narrowband drive sound source signal 24. As a method of determining the gain in the distortion minimizing means 23, a generally known Lagrangian undetermined coefficient method can be used.

  That is, the distortion minimizing means 23 receives the narrowband sound source signal 6 and the output of the narrowband adaptive codebook 22 as input, and has a distorted minimum lag length 25 and gain 26 which are narrowband adaptive sound source codes, and a narrowband driving sound source of the error signal. The signal 24 is output.

  The zero padding means 28 in the wideband driving sound source estimating means 27 inserts zero M-1 samples between each sample value of the narrowband driving sound source signal 24, and the obtained M times sample signal is the wideband driving sound source. Output as signal 29. Here, M is a value obtained by dividing the sampling frequency of the restored wideband audio signal by the sampling frequency of the narrowband audio signal, and the operation of inserting zero is the same as the zero padding means 15.

  In the wideband adaptive sound source estimation means 30, first, the narrowband adaptive lag length 25 is multiplied by M to obtain the wideband adaptive lag length 33, and the narrowband adaptive gain is multiplied by g to obtain the wideband adaptive gain. When g is 1, the pitch periodicity of the finally obtained wideband sound source signal 16 is equivalent to that of the narrowband sound source signal 6, and the pitch periodicity becomes weaker than that of the narrowband sound source signal 6 as the value is decreased from 1. Go. Observing actual speech, the higher the frequency, the more often the pitch periodicity becomes weaker. When restoring high frequencies, setting g to a value smaller than 1 can restore higher quality broadband speech. .

  The broadband adaptive excitation codebook 31 in the broadband adaptive excitation estimation means 30 stores the past broadband excitation signal 16 and outputs a signal obtained by repeating this signal with the broadband adaptive lag length 33. Then, this signal is multiplied by the broadband adaptive gain 34 in the broadband adaptive sound source estimation means 30 and output as a broadband adaptive sound source signal 32.

  Finally, the wide-band drive sound source signal 29 and the wide-band adaptive sound source signal 32 are added and output as the wide-band sound source signal 16.

  By configuring in this way, the characteristics related to the strength and fluctuation of the pitch periodicity of the narrowband sound source signal are well expressed by the narrowband adaptive lag length 25 and the narrowband adaptive gain 26, and reflected in the wideband sound source signal. Therefore, it is possible to sufficiently estimate variously changing sound sources, and there is an effect that it is possible to restore broadband sound with good sound quality without pulse-like sound. In addition, there is an effect that an appropriate sound source can be estimated regardless of the speaker.

  In the wideband adaptive sound source signal 32, the fundamental frequency determined by the wideband adaptive lag length 33 and the frequency of its harmonic component are correctly aligned at integer multiple positions, so that the narrowband component in the finally restored wideband audio signal 20 The restored wideband components are well connected, and high-quality wideband speech can be restored.

  Furthermore, since the feature that the pitch periodicity becomes weaker as the frequency becomes higher can be introduced by the coefficient g, there is an effect that a more natural sound quality can be obtained.

  In addition, voiced and unvoiced determination and pitch extraction are not necessary, and sound sources with intermediate properties can be expressed, so there is no influence of voiced and unvoiced determination errors and pitch extraction errors that tend to occur on narrowband audio signals with superimposed noise, A good broadband sound source can be estimated even near the voiced and unvoiced boundary, and there is an effect that it is possible to restore broadband sound with stable and natural sound quality.

Example 3
FIG. 5 is a configuration diagram of the wideband drive sound source estimation means 27 in the wideband speech restoration apparatus according to the third embodiment of the present invention. The new parts in the figure are 35 power calculation means and 36 noise generation means. Since other configurations are the same as those in FIGS. 1 and 3, the description of the operation of the corresponding parts is omitted.

  Hereinafter, the operation of the portion shown in FIG. 5 according to the third embodiment of the present invention will be described with reference to FIG.

  The narrow band driving sound source signal 24 is input to the power calculating means 35 in the wide band driving sound source estimating means 27. The power calculation means 35 calculates and outputs the power of the narrow band drive sound source signal 24. The noise generator 36 generates and outputs a power-normalized white noise signal. Then, in the wideband drive sound source estimation means 27, the white noise signal is multiplied by the power output from the power calculation means 35, and the obtained signal is output as the wideband drive sound source signal 29.

  The pitch period and the strength of periodicity change from moment to moment. Since the fine fluctuation of the pitch period and periodicity in the narrowband sound source signal 6 cannot be expressed by the narrowband adaptive lag length 25 and the narrowband adaptive gain 26, the error is included in the narrowband drive sound source signal 24. . When the wideband drive sound source signal 29 is generated using the narrowband drive sound source signal 24 including the error component as in the second embodiment, unnecessary disturbance may occur in the wideband drive sound source signal 29, and the power is the same. It has been experimentally confirmed that a better restored sound may be obtained when white noise is generated and used as the broadband driving sound source signal 29.

  By configuring as in the third embodiment, white noise having the same power as the narrow-band driving sound source signal 24 is generated and used as the wide-band driving sound source signal 29. Therefore, in addition to the effects of the second embodiment, the pitch period In addition, there is an effect of obtaining a good restoration sound with little disturbance due to fluctuations in the intensity of periodicity.

  When zero padding is performed, a symmetrical spectrum is generated around 4 kHz. Therefore, when zero padding is performed on the narrowband driving sound source signal 24 having no component of 0 Hz to 300 Hz and 3.4 KHz to 4.0 KHz, 0 Hz to 300 Hz, 3.4 KHz to 4.6 KHz, 7.7 KHz to 8 KHz. A signal having no component is obtained. On the other hand, in this configuration using white noise, the broadband driving sound source signal 29 having all components from 0 Hz to 8 KHz can be obtained, so that there is an effect that a good restoration sound having a band over the entire region can be obtained. The effect is particularly great when restoring from 0 Hz to 300 Hz.

Example 4
FIG. 6 is a configuration diagram of the wideband drive sound source estimation means 27 in the wideband speech restoration apparatus according to the fourth embodiment of the present invention. In the figure, 28 zero padding means, 35 power calculation means, and 36 noise generation means are the same as those in the second and third embodiments. Since other configurations are the same as those in FIGS. 1 and 3, description of operations other than those illustrated is omitted.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The narrow band drive sound source signal 24 is input to the zero padding means 28 and the power calculation means 35 in the wide band drive sound source estimation means 27. The zero padding means 28 in the wideband drive sound source estimation means 27 inserts zero M-1 samples between the sample values of the narrowband drive sound source signal 24, and outputs a signal having the number of samples obtained M times. Here, M is a value obtained by dividing the sampling frequency of the restored wideband audio signal by the sampling frequency of the narrowband audio signal, and the operation of inserting zero is the same as the zero padding means 15.

  The power calculation means 35 calculates and outputs the power of the narrow band drive sound source signal 24. The noise generator 36 generates and outputs a power-normalized white noise signal. Then, a signal obtained by multiplying the signal output by the zero padding means 28 by the gain gr1, the white noise signal output by the noise generating means 36 and the power output by the power calculating means 35, and a signal multiplied by the gain gr2 are added. And output as a broadband driving sound source signal 29.

  When the restored sounds according to the second and third embodiments have merits and demerits, respectively, by configuring in this way and setting gr1 and gr2 appropriately, it is possible to restore wideband speech with a quality exceeding both. effective. In addition, it has the same effect as Example 2 and Example 3.

Embodiment 5 FIG.
Another configuration capable of satisfactorily restoring the broadband sound source signal will be described.

  FIG. 7 is a block diagram of the broadband sound source estimation means 14 in the broadband speech restoration apparatus of Embodiment 5 of the present invention. The novel parts in the figure are: 37 narrowband long period prediction analysis means, 38 narrowband long period delay, 39 narrowband long period prediction coefficient, 40 long period inverse filter, 41 narrowband long period prediction residual Signal, 42 broadband long-period prediction residual estimation means, 43 zero padding means, 44 broadband long-period prediction parameter (code) estimation means, 45 broadband long-period delay, 46 broadband long-period prediction coefficient, 47 length It is a periodical synthesis filter, 48 wideband long period prediction residual signal. The overall configuration is the same as in FIG.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The narrow band sound source signal 6 is input to the sound source analysis means 21 in the wide band sound source estimation means 14. The narrowband long cycle prediction analysis unit 37 in the sound source analysis unit 21 performs long cycle prediction analysis on the narrowband sound source signal 6, and a narrowband long cycle delay 38 and a narrowband long cycle which are narrowband long cycle prediction codes. The prediction coefficient 39 is output. Note that the long-period prediction analysis is a method often used in the CELP encoding method, and thus the description thereof is omitted.

  The long-period inverse filter 40 in the sound source analysis means 21 performs inverse filtering on the narrow-band sound source signal 6 using the narrow-band long-period delay 38 and the narrow-band long-period prediction coefficient 39, and the obtained signal is subjected to narrow-band long-period prediction. The residual signal 41 is output to the wideband long period prediction residual estimation means 42.

  The zero padding means 43 in the wideband long period prediction residual estimation means 42 inserts zero M-1 samples between each sample value of the narrowband long period prediction residual signal 41, and M times the number of samples obtained. The signal is output as a wideband long period prediction residual signal 48. Here, M is a value obtained by dividing the sampling frequency of the restored wideband audio signal by the sampling frequency of the narrowband audio signal, and the operation of inserting zero is the same as the zero padding means 15.

  The wideband long period prediction parameter (code) estimation means 44 multiplies the narrowband long period delay 38 by M and outputs a wideband long period delay 45 which is one of the prediction codes, and also sets the narrowband long period prediction coefficient 39 as g The wideband long-period prediction coefficient 46, which is another prediction code, is output. When g is 1, the pitch periodicity of the finally obtained wideband sound source signal 16 is equivalent to that of the narrowband sound source signal 6, and the pitch periodicity becomes weaker than that of the narrowband sound source signal 6 as the value is decreased from 1. Go. Similar to the second embodiment, when restoring a high frequency band, setting g to a value smaller than 1 results in higher quality.

  Finally, the long-period synthesis filter 47 performs long-period synthesis filtering on the wide-band long-period prediction residual signal 48 using the wide-band long-period delay 45 and the wide-band long-period prediction coefficient 46, and the obtained signal is wide-band The sound source signal 16 is output.

  By configuring in this way, characteristics regarding the strength and fluctuation of pitch periodicity possessed by the narrowband sound source signal are well expressed by the narrowband long period delay 38 and the narrowband long period prediction coefficient 39 and reflected in the wideband sound source signal. Therefore, it is possible to sufficiently estimate variously changing sound sources, and to restore a wide-band sound having a good sound quality without a pulse sound. In addition, there is an effect that an appropriate sound source can be estimated regardless of the speaker.

  In the broadband sound source signal 16, the fundamental frequency determined by the broadband long-cycle delay 45 and the frequency of its harmonic component are correctly aligned at integer multiple positions, so that the narrowband component and the reconstruction in the finally restored broadband audio signal 20 are restored. The broadband components are well connected, and there is an effect that high-quality broadband speech can be restored.

  Furthermore, since the feature that the pitch periodicity becomes weaker as the frequency becomes higher can be introduced by the coefficient g, there is an effect that a more natural sound quality can be obtained.

  In addition, voiced and unvoiced determination and pitch extraction are not necessary, and sound sources with intermediate properties can be expressed, so there is no influence of voiced and unvoiced determination errors and pitch extraction errors that tend to occur on narrowband audio signals with superimposed noise, A good broadband sound source can be estimated even near the voiced and unvoiced boundary, and there is an effect that it is possible to restore broadband sound with stable and natural sound quality.

Example 6
FIG. 8 is a block diagram of the wideband long period prediction residual estimation means 42 in the wideband speech restoration apparatus of Embodiment 6 of the present invention. In the figure, 35 power calculation means and 36 noise generation means are the same as those in the third embodiment. Other configurations are the same as those in FIG. 1 and FIG.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  The narrowband long period prediction residual signal 41 is input to the power calculation means 35 in the wideband long period prediction residual estimation means 42. The power calculation means 35 calculates and outputs the power of the narrowband long period prediction residual signal 41. The noise generator 36 generates and outputs a power-normalized white noise signal. Then, in the wideband long period prediction residual estimation means 42, the white noise signal is multiplied by the power output from the power calculation means 35, and the obtained signal is output as a wideband long period prediction residual signal 48.

  Similar to the description in the third embodiment, the minute fluctuations in the pitch period and periodicity in the narrowband sound source signal 6 cannot be expressed by the narrowband long period delay 38 and the narrowband long period prediction coefficient 39, and therefore the error. Is included in the narrowband long period prediction residual signal 41. When the wideband long period prediction residual signal 48 is generated using the narrowband long period prediction residual signal 41 including the error component as in the fifth embodiment, unnecessary disturbance occurs in the wideband long period prediction residual signal 48. In some cases, it is possible to obtain a better restored sound when white noise having the same power is generated and used as the wideband long-period prediction residual signal 48.

  By configuring as in the sixth embodiment, white noise having the same power as that of the narrowband long-period prediction residual signal 41 is generated and used as the wideband long-period prediction residual signal 48. In addition to this, there is an effect that a good restoration sound with little disturbance due to the fluctuation of the pitch period and the intensity of periodicity can be obtained.

  Further, when zero padding is performed, a symmetric spectrum is generated around 4 KHz. Therefore, this is performed for the narrowband long period prediction residual signal 41 having no components from 0 Hz to 300 Hz and 3.4 KHz to 4.0 KHz. Then, a signal having no component of 0 Hz to 300 Hz, 3.4 KHz to 4.6, and KHz 7.7 KHz to 8 KHz is obtained. On the other hand, in this configuration using white noise, the wideband long-period prediction residual signal 48 having all components from 0 Hz to 8 KHz can be obtained, so that there is an effect that a good restored sound with no insufficient band can be obtained. The effect is particularly great when restoring from 0 Hz to 300 Hz.

Example 7
FIG. 9 is a configuration diagram of the wideband long-period prediction residual estimation means 42 in the wideband speech restoration apparatus according to the seventh embodiment of the present invention. In the figure, 43 zero-filling means, 35 power calculating means, and 36 noise generating means are the same as those in the fifth and sixth embodiments. Other configurations are the same as those in FIG. 1 and FIG.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  The narrowband long period prediction residual signal 41 is input to the zero padding means 43 and the power calculation means 35 in the wideband long period prediction residual estimation means 42. The zero padding means 43 in the wideband long period prediction residual estimation means 42 inserts zero M-1 samples between each sample value of the narrowband long period prediction residual signal 41, and M times the number of samples obtained. The signal is output. Here, M is a value obtained by dividing the sampling frequency of the restored wideband audio signal by the sampling frequency of the narrowband audio signal, and the operation of inserting zero is the same as the zero padding means 15.

  The power calculation means 35 calculates and outputs the power of the narrowband long period prediction residual signal 41. The noise generator 36 generates and outputs a power-normalized white noise signal. Then, a signal obtained by multiplying the signal output by the zero padding means 43 by the gain gr1, the white noise signal output by the noise generating means 36 and the power output by the power calculating means 35, and a signal multiplied by the gain gr2 are added. And output as a wideband long-period prediction residual signal 48.

  In the case where the restored sounds according to the fifth and sixth embodiments have advantages and disadvantages, respectively, it is possible to recover wideband speech with a quality exceeding both by configuring in this manner and appropriately setting gr1 and gr2. effective. In addition, it has the same effect as Example 5 and Example 6.

Example 8 FIG.
FIG. 10 is a configuration diagram of the wideband sound source estimation means 14 in the wideband speech restoration apparatus according to the eighth embodiment of the present invention. The novel parts in the figure are 49 upsampling means and 50 nulling means. The overall configuration is the same as in FIG.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  The narrow band sound source signal 6 is input to the upsampling means 49. The upsampling means 49 upsamples the narrow band sound source signal 6 M times and outputs the obtained signal to the sound source analysis means 21.

  The narrowband long cycle prediction analysis unit 37 in the sound source analysis unit 21 performs long cycle prediction analysis on the output signal of the upsampling unit 49 and outputs a narrowband long cycle delay 38 and a narrowband long cycle prediction coefficient 39. . Note that the delay search range in the long-period prediction analysis is M times that in the fifth embodiment.

  The long-period inverse filter 40 in the sound source analysis means 21 uses the narrow-band long-period delay 38 and the narrow-band long-period prediction coefficient 39 to inverse-filter the output signal of the up-sampling means 49, and the obtained signal is narrow-band. The long-period prediction residual signal 41 is output to the broadband long-period prediction residual estimation means 42.

  The zeroing means 50 in the wideband long period prediction residual estimation means 42 leaves only the signal of every M samples of the narrowband long period prediction residual signal 41 and sets the value of the remaining signal to zero. Then, the obtained signal is output as a wideband long period prediction residual signal 48.

  The wideband long period prediction parameter estimation means 44 outputs the narrowband long period delay 38 as it is as the wideband long period delay 45, multiplies the narrowband long period prediction coefficient 39 by g, and outputs it as a wideband long period prediction coefficient 46. About g, it is the same as that of Example 5.

  Finally, the long-period synthesis filter 47 performs long-period synthesis filtering on the wide-band long-period prediction residual signal 48 using the wide-band long-period delay 45 and the wide-band long-period prediction coefficient 46, and the obtained signal is wide-band The sound source signal 16 is output.

  By configuring in this way, it is possible to perform long-period analysis on signals with a high sampling frequency, so that delays with higher accuracy can be analyzed, and characteristics regarding the strength and fluctuation of pitch periodicity of narrowband sound source signals Can be reflected more finely in the broadband sound source signal, and it is possible to sufficiently estimate variously changing sound sources and to restore broadband sound with good sound quality. It has the same effect as the fifth embodiment.

Example 9
FIG. 11 is a configuration diagram of the wideband speech restoration apparatus according to the ninth embodiment of the present invention. The novel parts in the figure are 51 narrowband power calculation means, 52 narrowband excitation power, and 53 narrowband power-included spectrum codebook. Others are the same as those described above, so only those having slight differences in operation will be described.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The narrowband power calculation means 51 in the analysis means 2 calculates the power included in the amplitude information of the narrowband sound source signal 6 and outputs it as the narrowband sound source power 52. In addition, the spectral parameter 4 and the narrowband sound source signal 6 are also output.

  The vector quantization means 8 in the wideband spectrum estimation means 7 uses the narrowband power-included spectrum codebook 53 to vector quantize the narrowband spectrum parameter 4 and the narrowband excitation power 52 in a lump, and obtain the obtained spectrum code. 10 is output to the inverse quantization means 11 in the broadband spectrum estimation means 7.

  Here, the narrowband power-included spectrum codebook 53 is created in the same manner as in Document 1 using a pair of narrowband spectral parameters and narrowband excitation power obtained by analyzing many narrowband speech signals as learning data. To do. As a distance scale in the learning of the narrowband power-included spectrum codebook 53 and the vector quantization means 8, a value obtained by multiplying the Euclidean distance of the logarithmic value of power by w times and the Euclidean distance of the spectrum parameter can be used. .

  Note that the narrowband power calculation means 51 can calculate the power of the narrowband sound signal 1 instead of the narrowband sound source signal 6 and use it instead of the narrowband sound source power 52. In this case, the narrowband power-included spectrum codebook 53 is learned using a pair of narrowband spectrum parameters and narrowband speech signal power as learning data.

  By configuring in this way, in addition to the effects of the first embodiment, information on power is reflected in the estimation of broadband spectral parameters, and there is an effect that a good spectrum can be estimated more stably.

Example 10
FIG. 12 is a configuration diagram of the wideband speech restoration apparatus according to the tenth embodiment of the present invention. The novel parts in the figure are included in 54 sound source normalizing means, 55 narrowband normalized sound source signal, 56 wideband normalized sound source signal, 57 wideband power codebook, 58 wideband sound source power, and wideband spectrum estimating means. 59 broadband sound source power estimation means. Since others are the same as those described above, the description thereof is omitted.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The sound source normalization means 54 in the analysis means 2 calculates the power contained in the amplitude information of the narrow band sound source signal 6 and outputs it as the narrow band sound source power 52 to the wide band sound source power estimation means 59 and also the narrow band sound source signal 6. Is output to the wideband sound source estimation means 14 as a narrowband normalized sound source signal 55.

  Actually, the vector quantization means 8 in the wideband excitation power estimation means 59 in the wideband spectrum estimation means 7 uses the narrowband power-included spectrum codebook 53 to obtain the narrowband spectrum parameter 4 and the narrowband excitation power 52. Vector quantization is performed collectively, and the obtained spectrum code 10 is output to the inverse quantization means 11 in the broadband sound source power estimation means 59. The inverse quantization means 11 decodes the spectrum code 10 using the wide band power codebook 57 and outputs the obtained wide band excitation power 58.

  The wideband sound source estimation unit 14 estimates the wideband normalized sound source signal 56 using the narrowband normalized sound source signal 54. For the estimation in the broadband spectrum estimation means 7 and the broadband sound source estimation means 14, the same method as in the first to eighth embodiments can be used. Then, the broadband normalized sound source signal 56 is multiplied by the broadband sound source power 58 to generate the broadband sound source signal 16.

  By configuring in this way, in addition to the effect of the first embodiment, it is possible to reflect the difference in the spectrum parameters in the estimation of the broadband sound source power, so that it is possible to restore the broadband sound having a more correct amplitude. .

Example 11
FIG. 13 is a configuration diagram of the wideband speech restoration apparatus according to the eleventh embodiment of the present invention. The novel part in the figure is a 60 wideband power-included spectrum codebook. Others are the same as those in FIG. 11 and FIG. 12, and therefore only those having a slight difference in operation will be described.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The inverse quantization means 11 in the wideband spectrum estimation means 7 decodes the spectrum code 10 using the wideband power-included spectrum codebook 60 and outputs the obtained wideband spectrum parameter 13 and the wideband sound source power 58.

  Here, the broadband power-included spectrum codebook 60 is created in the same manner as in Document 1 using a pair of broadband spectrum parameters and broadband sound source power obtained by analyzing many broadband speech signals as learning data. The distance scale is the same as that used to create the narrowband power-included spectrum codebook 53.

  By configuring in this way, it is possible to combine the effects of the ninth and tenth embodiments.

Example 12 FIG.
FIG. 14 is a configuration diagram of the wideband speech restoration apparatus according to the twelfth embodiment of the present invention. The novel part in the figure is 61 post filter means. Others are the same as those of the first to eleventh embodiments, and the description thereof is omitted.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The post filter means 61 performs post filtering processing on the provisional broadband audio signal output from the synthesis filter 17 and outputs the obtained signal to the band filter 18. The band filter 18 performs a band pass filter process on the signal output from the post filter unit 61 to extract components other than the band having the narrow band audio component.

  The post-filtering process is a signal processing process that improves auditory quality. It emphasizes the pitch periodicity and spectral poles, emphasizes high frequencies to improve clarity, and passes through the transmission line. This suppresses a band having a large amount of distortion that occurs in the image to reduce the sense of distortion.

  As pitch pitch emphasis processing, a general method is to multiply the provisional wideband audio signal preceding the pitch period by a coefficient smaller than 1 and add it to the current provisional wideband audio signal.

  As pole emphasis processing, the broadband spectrum parameter 13 is modified to have a large gain in the frequency band near the pole frequency of the broadband spectrum parameter 13 and a small gain in a frequency band other than the pole vicinity of the broadband spectrum parameter 13. Various methods for calculating the filter coefficient of a pole-zero filter have been proposed and can be realized by applying this filter to a provisional wideband audio signal. In addition, since distortion generated when passing through the transmission line is large in a frequency band with a small amplitude, that is, in a frequency band other than the vicinity of the pole, a band with much distortion can be suppressed by this pole enhancement processing.

  As the high-frequency emphasis processing, a method called pre-emphasis, that is, a method of multiplying a temporary broadband audio signal one point before by a coefficient of 1 or less and subtracting it from the current temporary broadband audio signal is common.

  In FIG. 14, the post filter unit 61 and the band filter 18 may be in opposite positions, or the post filter unit 61 may be applied to the wideband audio signal 20.

  By configuring in this way, in addition to the effects of the first embodiment, when the sound quality of the restored wideband audio signal is insufficient, the pitch periodicity and spectrum pole of the wideband audio signal are emphasized, By emphasizing the above, it is possible to improve the clarity, and to suppress a band having a lot of distortion that occurs when passing through the transmission path, thereby reducing the distortion.

  In FIG. 14, a configuration in which the inverse filter 5 and the broadband sound source estimation unit 14 are removed is also possible. This configuration corresponds to the application of the present invention to Document 1, and has the same effect as described above.

Example 13
A configuration in which the broadband spectrum estimation means 7 in the first to twelfth embodiments outputs the narrowband spectrum parameter 4 as it is as the broadband spectrum parameter 13 is also possible.

  FIG. 15 is an explanatory diagram for explaining the relationship between the narrowband spectrum and the broad spectrum in this case. When the spectral envelope represented by the narrowband spectral parameter 4 is as shown in FIG. 15 (a), if this is used as it is as the wideband spectral parameter 13, the resulting width expands, and the spectral envelope represented by the wideband spectral parameter 13 is as shown in FIG. When (a) is extended M times in the frequency axis direction and M is 2, the result is as shown in FIG. Therefore, when the narrow-band spectrum envelope is high from 2 KHz to 3.4 KHz, the restored high frequency is higher than 3.4 KHz, and conversely, when 2 KHz to 3.4 KHz is low, the high frequency is low. As a result, the rough slope of the narrow-band spectral envelope is directly reflected in the high band.

  By configuring in this way, in addition to the effect of the first embodiment, there is an effect that the broadband spectrum can be reconstructed very easily, though roughly. Compared to the first embodiment, there is no need for a memory for storing a codebook, and the amount of calculation is reduced.

Example 14
In the first to twelfth embodiments, the wideband spectrum estimation unit 7 may output the narrowband spectrum parameter 4 from the lowest order to the predetermined order as the wideband spectrum parameter 13. However, the narrowband spectrum parameter 4 output from the spectrum analysis means 3 is always stable even if the band spectrum parameter 13 obtained from the lowest order to a predetermined order such as a PARCOR coefficient or autocorrelation coefficient is used. Only if the parameter is

  FIG. 16 is an explanatory diagram for explaining the relationship between the narrowband spectrum and the broad spectrum in this case. When the spectrum envelope represented by the narrowband spectrum parameter 4 is FIG. 16A, when the spectrum order from the lowest order to the predetermined order is used as the broadband spectrum parameter 13, the spectrum envelope represented by the broadband spectrum parameter 13 is as shown in FIG. Is stretched M times in the frequency axis direction to further smooth the polar structure. When M is 2, the result is as shown in FIG. As a result, it is possible to suppress the rough inclination of the narrow-band spectrum envelope as it is in the high band and the generation of a strong pole that does not exist in the high band, thereby generating an unnatural restoration sound.

  By configuring in this way, in addition to the effects of the thirteenth embodiment, in the rare case of the thirteenth embodiment, a strong non-existent pole is generated in the high range, thereby suppressing the occurrence of unnatural restoration sound. There is an effect that can be done.

Example 15.
FIG. 17 is a configuration diagram of the wideband spectrum estimation means 7 of the wideband speech restoration apparatus according to the fifteenth embodiment of the present invention. The novel parts in the figure are 62 spectral parameter conversion means, 63 order reduction means, and 64 spectral parameter inverse conversion means. Others are the same as those of the first to twelfth embodiments, and the description thereof is omitted.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  The spectral parameter conversion means 62 in the wideband spectrum estimation means 7 converts the narrowband spectral parameter 4 into a parameter whose composition is always stable when the lowest order to a predetermined order such as PARCOR coefficient and autocorrelation coefficient are extracted. To do. The order reduction unit 63 outputs the parameter output from the lowest order to the predetermined order output by the spectral parameter conversion unit 62 to the spectral parameter inverse conversion unit 64. The spectrum parameter inverse conversion means 64 returns the parameter output from the order reduction means 63 to the same region as the narrowband spectrum parameter 4 and outputs it as the broadband spectrum parameter 13.

  By configuring in this way, the same effect as in the fourteenth embodiment can be obtained even when the narrowband spectrum parameter 4 is a parameter in which the synthesis becomes unstable when the lowest order to the predetermined order are extracted.

Example 16
In the fourteenth and fifteenth embodiments, strong poles are suppressed by reducing the order, but a method that gives a similar effect, such as using an autocorrelation coefficient as a spectral parameter and applying a lag window thereto, can be used.

  By configuring in this way, the same effect as that of the embodiment 14 can be obtained by another means.

  Note that the broadband spectrum estimation means 7 of the thirteenth to sixteenth embodiments can be applied to a conventional configuration such as Document 1. For example, the overall configuration when applied to Document 1 is obtained by removing the inverse filter 5, the broadband sound source estimation unit 14, and the post filter unit 61 from FIG. 14. In such a configuration, the effects newly generated in the embodiments 13 to 16 can be added to the conventional technology.

Example 17.
In the following embodiment, an example will be described in which the present invention is applied to an apparatus that restores wideband speech based on encoded information by transmission or the like.

  FIG. 18 is a block diagram of a wideband speech restoration apparatus according to Embodiment 17 of the present invention. In the figure, 101 is a narrowband speech code, 102 is a separation means, 103 is a narrowband spectrum code, 104 is a narrowband excitation code, 105 is a wideband spectrum decoding means, 106 is a wideband excitation decoding means, and 107 is a narrowband spectrum decoding means. , 108 are narrowband sound source decoding means, and 109 is narrowband speech decoding means. Others are the same as those of the first to sixteenth embodiments, and the description thereof is omitted.

  Also in the present embodiment, a good broadband sound source signal is obtained without performing reanalysis.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  First, the narrowband speech code 101 is input to the separating unit 102 and the narrowband speech decoding unit 109. The narrowband speech code 101 is a separately encoded narrowband speech signal sampled at, for example, 8 KHz and limited to a telephone band from 300 Hz to 3.4 KHz, and is input from a storage medium or a communication channel. Is.

  Separating means 102 separates narrowband speech code 101 into narrowband spectral code 103 and narrowband excitation code 104, narrowband spectral code 103 into wideband spectrum decoding means 105, and narrowband excitation code 104 into wideband excitation codec means. The data is output to 106.

  The narrowband spectrum decoding means 107 in the wideband spectrum decoding means 105 decodes the narrowband spectrum code 103 and outputs the obtained narrowband spectrum parameter 4. Note that the narrowband spectrum decoding means 107 may perform the reverse process of the encoding process of the narrowband spectrum parameter used when the narrowband speech code 101 is encoded.

  Then, the broadband spectrum estimation means 7 in the broadband spectrum decoding means 105 estimates the broadband spectrum parameter 13 using the narrowband spectrum parameter 4. As the broadband spectrum estimation means 7, the methods described in the embodiments described so far can be used.

  The narrowband excitation decoding means 108 in the wideband excitation decoding means 106 decodes the narrowband excitation code 104 and outputs the obtained narrowband excitation signal 6. Then, the broadband excitation estimation means 14 in the broadband excitation decoding means 106 estimates the broadband excitation signal 16 using the narrowband excitation signal 6.

  The broadband sound source estimation means 14 can be zero padding means or the like. The narrowband excitation decoding means 108 may perform the reverse process of the encoding process of the narrowband excitation signal used when the narrowband speech code 101 is encoded.

  The synthesis filter 17 performs synthesis filter processing on the broadband sound source signal 16 using the broadband spectrum parameter 13 to generate a temporary broadband audio signal. The band filter 18 performs band pass filter processing on the provisional wideband audio signal, and extracts components other than the band having the narrowband audio component. When the band of the wideband audio signal is from 0 Hz to 7.3 KHz, components from 0 Hz to 300 Hz and 3.4 KHz to 7.3 KHz are extracted.

  On the other hand, the narrowband speech decoding unit 109 decodes the input narrowband speech code 101 and outputs the obtained narrowband speech signal 1 to the upsampling unit 19. This decoding process may be performed by reversing the encoding process used when the narrowband speech code 101 is encoded.

  Next, the upsampling means 19 upsamples the narrowband audio signal 1 M times. The signal generated by the upsampling has the same narrowband component as the narrowband audio signal 1 with the same sampling frequency as the wideband audio signal 20. Then, the output of the band filter 18 and the output of the upsampling means 19 are added to generate a wideband audio signal 20.

  With this configuration, when a narrowband speech code is received from a storage medium or a communication channel, there is no need to reanalyze the narrowband speech, so that it can be restored with a small amount of processing. Further, since distortion due to interpolation at the time of synthesis or windowing at the time of analysis is not superimposed, there is an effect that a broadband voice with better quality can be restored. It has the same effect as the first embodiment.

  The narrowband speech decoding unit 109 may be configured to synthesize the narrowband speech signal 1 by inputting the narrowband spectrum parameter 4 and the narrowband excitation signal 6, or conversely, the decoding in the narrowband speech decoding unit 109. A configuration in which the narrowband spectral parameter 4 and the narrowband excitation signal 6 calculated as intermediate parameters of the process are input to the wideband spectrum decoding means 105 and the wideband excitation decoding means 106 is also possible. In this case, it is possible to omit the overlapping processing and to recover the wideband sound with a smaller processing amount.

  When the pitch period code and the power code can be separated from the narrowband speech code 101, the pitch period and power information are decoded from these codes, and the wideband spectrum parameter 13 and the pitch period and power information are used. Thus, it is possible to generate a temporary broadband synthesized sound by the same method as in Document 1.

Example 18
FIG. 19 is a block diagram of the wideband excitation decoding means 106 of the wideband speech restoration apparatus according to the eighteenth embodiment of the present invention. The novel parts in the figure are 110 narrowband pitch code, 111 narrowband power code, 112 wideband pitch decoding means, 113 wideband pitch period, 114 wideband power decoding means, 115 wideband power, and 116 sound source generation. Means. Others are the same as those of the embodiment 17, and the description is omitted.

  This embodiment is limited to the case where a narrowband speech code 101 that can easily separate the narrowband pitch code 110 and the narrowband power code 111 by the separating means 102 is input. In this case, the configuration of FIG. 19 is meaningful.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  As the narrowband excitation code 104, a narrowband pitch code 110 and a narrowband power code 111 are input to the wideband excitation decoding means 106.

  Wideband pitch decoding means 112 in wideband excitation decoding means 106 estimates wideband pitch period 113 using narrowband pitch code 110. As an estimation method, the narrowband pitch period may be decoded from the narrowband pitch code 110 and the value may be multiplied by M, but the result is held as a table and the table component corresponding to the narrowband pitch code 110 is stored. You may ask by reading.

  Next, the broadband power decoding unit 114 in the broadband excitation decoding unit 106 estimates the broadband power 115 using the narrowband power code 111. As an estimation method, the narrowband power may be decoded from the narrowband power code 111 and the value may be multiplied by g. However, a table component corresponding to the narrowband power code 111 is obtained by holding the result as a table. You may ask for it by reading.

  The sound source generating means 116 outputs a signal in which fixed sound sources are arranged with the wide band pitch period 113 as a repetition period, and finally multiplies the output signal of the sound source generating means 116 by the wide band power 115 and outputs it as the wide band sound source signal 16. To do.

  By configuring in this way, in addition to the effect of the seventeenth embodiment, the wideband sound source signal 16 is directly generated without decoding the narrowband sound source signal, so that there is an effect that restoration can be performed with a small amount of processing.

Example 19.
FIG. 20 is a configuration diagram of the wideband sound source decoding means 106 of the wideband speech restoration apparatus according to the nineteenth embodiment of the present invention. The novel parts in the figure are 117 narrowband adaptive excitation codes, 118 narrowband drive excitation codes, 119 wideband adaptive excitation decoding means, 120 wideband excitation excitation decoding means, 121 narrowband adaptive excitation decoding means, This is narrowband drive excitation decoding means. Others are the same as described above, and a description thereof will be omitted.

  This embodiment is limited to the case where a narrowband speech code 101 that allows the narrowband adaptive excitation code 117 and the narrowband drive excitation code 118 to be easily separated from the input narrowband speech code by the separation means 102 is input. It is done. In this case, the configuration of FIG. 20 is meaningful.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  As the narrowband excitation code 104, a narrowband adaptive excitation code 117 and a narrowband drive excitation code 118 are input to the wideband excitation decoding means 106.

  The narrowband adaptive excitation decoding means 121 in the wideband adaptive excitation decoding means 119 decodes the narrowband adaptive excitation code 117 and outputs the obtained narrowband adaptive lag length 25 and narrowband adaptive gain 26. Wideband adaptive excitation estimation means 30 in wideband adaptive excitation decoding means 119 generates and outputs wideband adaptive excitation signal 32 from narrowband adaptive lag length 25 and narrowband adaptive gain 26. The operation of the broadband adaptive sound source estimation means 30 is the same as that of the second embodiment.

  The narrowband drive excitation decoding means 122 in the wideband drive excitation decoding means 120 decodes the narrowband drive excitation code 118 and outputs the obtained narrowband drive excitation signal 24. The broadband drive excitation estimation means 27 in the broadband drive excitation decoding means 120 estimates and outputs a broadband drive excitation signal 29 from the narrowband drive excitation signal 24. The operation of the wideband drive sound source estimation means 27 is the same as in the second to fourth embodiments.

  Finally, the wideband adaptive sound source signal 32 and the wideband drive sound source signal 29 are added and output as the wideband sound source signal 16.

  With this configuration, in addition to the effects of the second to fourth embodiments and the seventeenth embodiment, the wideband sound source signal 16 is directly generated without decoding the narrowband sound source signal. There is an effect that can be restored.

  In addition, since the fundamental frequency and the frequency of its harmonic components are correctly aligned at integer multiples, the narrow-band component and the restored broadband component in the finally restored broadband audio signal are well connected, resulting in high-quality broadband audio. There is an effect that can be restored.

  In addition, since voiced unvoiced information and pitch period information are not used, it is possible to express sound sources with intermediate characteristics, so that the effects of voiced unvoiced decision errors and pitch extraction errors that tend to occur on narrowband audio signals with superimposed noise are affected. In addition, a good broadband sound source can be estimated even near the voiced and unvoiced boundary, and stable and natural sound quality broadband sound can be restored.

Example 20.
FIG. 21 is a configuration diagram of the wideband excitation decoding means 106 of the wideband speech restoration apparatus according to the twentieth embodiment of the present invention. In the figure, the new parts are 123 narrowband long period prediction code, 124 wideband long period prediction parameter (code) decoding means, 125 narrowband long period prediction parameter (code) decoding means, 126 narrowband long period prediction code. They are a residual code, 127 wide-band long-period prediction residual decoding means, and 128 narrow-band long-period prediction residual decoding means. Others are the same as described above, and a description thereof will be omitted.

  In this embodiment, a narrowband speech code 101 that can easily separate a narrowband long period prediction code 123 and a narrowband long period prediction residual code 126 from an input narrowband speech code by the separation means 102 is input. Limited to In this case, the configuration of FIG. 21 is meaningful.

  Hereinafter, the operation of the embodiment of the present invention will be described with reference to FIG.

  As the narrowband excitation code 104, a narrowband long period prediction code 123 and a narrowband long period prediction residual code 126 are input to the wideband excitation decoding means 106.

  The narrowband long period prediction parameter decoding means 125 in the wideband long period prediction parameter (code) decoding means 124 decodes the narrowband long period prediction code 123 and is a narrowband length which is one of the obtained prediction codes. A period delay 38 and a narrowband long period prediction coefficient 39 which is another prediction code are output. The wideband long period prediction parameter estimation unit 44 in the wideband long period prediction parameter decoding unit 124 uses the narrowband long period delay coefficient 38 and the narrowband long period prediction coefficient 39 to generate a wideband long period delay which is one of the long period prediction codes. 45 and a wideband long-period prediction coefficient 46, which is one of the other long-period prediction codes, are estimated and output. The operation of the wideband long period prediction parameter estimation unit 44 is the same as that of the fifth embodiment.

  The narrowband long period prediction residual decoding means 128 in the wideband long period prediction residual decoding means 127 decodes the narrowband long period prediction residual code 126 and obtains the narrowband long period prediction residual signal 41 obtained. Is output. The wideband long period prediction residual estimation means 42 in the wideband long period prediction residual decoding means 127 estimates the wideband long period prediction residual signal 48 from the narrowband long period prediction residual signal 41 and outputs it. The operation of the wideband long period prediction residual estimation means 42 is the same as in the fifth to seventh embodiments.

  Finally, the long-period synthesis filter 47 performs long-period synthesis filtering on the wide-band long-period prediction residual signal 48 using the wide-band long-period delay 45 and the wide-band long-period prediction coefficient 46, and the obtained signal is wide-band The sound source signal 16 is output.

  With this configuration, in addition to the effects of the fifth to seventh embodiments and the seventeenth embodiment, the wideband sound source signal 16 is directly generated without decoding the narrowband sound source signal. There is an effect that can be restored.

Example 21.
In the seventeenth to the twentieth embodiments, the wideband spectrum parameter 13 is estimated after the narrowband spectrum code 4 is decoded from the narrowband spectrum code 103. The narrowband spectrum code 103 refers to the wideband spectrum codebook. A configuration in which the broadband spectral parameter 13 is directly calculated is also possible.

  By configuring in this way, in addition to the effects of Embodiments 17 to 20, there is an effect that restoration can be performed with a smaller amount of processing.

Example 22.
FIG. 22 is a block diagram of a wideband speech restoration apparatus according to an embodiment of the present invention. The novel parts in the figure are 129 narrowband power decoding means and 130 wideband normalized excitation decoding means.

  The broadband spectrum estimation means 7 is the same as that of the eleventh embodiment, and the others are the same as those described above, and the description thereof is omitted.

  The operation of one embodiment of the present invention will be described below with reference to FIG.

  The narrowband power decoding unit 129 decodes a portion related to power from the narrowband amplitude information included in the narrowband excitation code 104 and outputs the obtained narrowband excitation power 52 to the wideband spectrum estimation unit 7. . The wideband spectrum estimation means 7 estimates the wideband spectrum parameter 13 and the wideband sound source power 58 using the narrowband spectrum parameter 4 and the narrowband sound source power 52.

  The wideband normalized excitation decoding means 130 estimates a wideband excitation signal whose power is normalized by using a portion other than the portion related to the narrowband power included in the narrowband excitation code 104, as a wideband normalized excitation signal 56. Output. For the processing in the wideband normalized excitation decoding means 130, the same processing as that in the eighteenth to twentieth embodiments can be used. Then, the broadband normalized sound source signal 56 is multiplied by the broadband sound source power 58 to generate the broadband sound source signal 16.

  By comprising in this way, the effect which Example 11 and Example 18 thru | or Example 20 have can be combined. A configuration in which the broadband spectrum estimation means 7 estimates only one of the broadband spectrum parameter 13 or the broadband sound source power 58 as in the ninth and tenth embodiments is also possible.

Example 23.
In Embodiments 17 to 22, a configuration in which the post filter means 61 is inserted between the synthesis filter 17 and the band filter 18 is also possible. Further, the post filter unit 61 and the band filter 18 can be arranged in the reverse positions, and the post filter unit 61 can be applied to the wideband audio signal 20.

  With this configuration, when post-filter processing is performed in the narrowband speech decoding means 109, continuity between the narrowband portion and the restored band can be improved. Further, the effects of the embodiment 12 and the embodiments 17 to 22 can be combined.

Example 24.
In the configuration in which the wide band excitation decoding means 106 is removed from FIG. 18, a configuration in which the post filter means 61 is inserted between the synthesis filter 17 and the band filter 18 is also possible. Further, the post filter unit 61 and the band filter 18 can be arranged in the reverse positions, and the post filter unit 61 can be applied to the wideband audio signal 20.

  This configuration corresponds to that obtained by applying the twelfth embodiment to the seventeenth embodiment of the present invention in document 1, and when post-filter processing is performed in the narrowband speech decoding means 109, the narrowband portion and the restored band It has the effect of improving continuity.

  Below, the characteristics of each embodiment will be described together.

  The above-described wideband speech restoration apparatus includes an analysis unit that analyzes a narrowband speech signal to obtain a narrowband spectrum parameter and a narrowband sound source signal, a spectrum estimation unit that estimates a broadband spectrum parameter using the narrowband spectrum parameter, Broadband sound source estimation means for estimating a wideband sound source signal using a narrowband sound source signal and synthesis means for generating a wideband audio signal from the estimated wideband spectral parameters and the wideband sound source signal are provided.

  Further, as the wideband sound source estimation means, zero padding means for inserting a predetermined zero value into each sample interval of the input narrowband sound source signal is used.

  The wideband excitation estimating means analyzes the input narrowband excitation signal to obtain a narrowband adaptive excitation code and a narrowband drive excitation signal, and a wideband adaptive excitation signal using the narrowband adaptive excitation code. Adaptive sound source estimating means for estimating, driving sound source estimating means for estimating a wide-band driving sound source signal using a narrow-band driving sound source signal, and generating a wide-band sound source signal from the estimated wide-band adaptive sound source signal and wide-band driving sound source signal It comprised with the addition means.

  Alternatively, the wideband sound source estimation means analyzes the input narrowband sound source signal to obtain a narrowband long period prediction code and a narrowband long period prediction residual signal, and the narrowband long period prediction residual signal. A long-period prediction residual estimation means for estimating a wideband long-period prediction residual signal using a wideband long-period prediction code estimation means for estimating a wideband long-period prediction code using a narrowband long-period prediction code; A long-period synthesis means for synthesizing a wide-band source signal from a wide-band long-period prediction residual signal and a wide-band long-period prediction code.

  Another wideband speech restoration apparatus analyzes narrowband speech signals to obtain narrowband spectral parameters and narrowband amplitude information, and uses at least the wideband spectral parameters or narrowband amplitude information using the narrowband spectral parameters and narrowband amplitude information. Spectral estimation means for estimating wideband amplitude information and synthesis means for generating a wideband speech signal from these estimated wideband spectral parameters and wideband amplitude information or wideband sound source signals are provided.

  Alternatively, a wideband estimation unit that estimates a wideband audio signal using a narrowband audio signal and a postfilter unit that performs postfiltering on the estimated wideband audio signal are provided.

  Or analyzing means for analyzing a narrowband speech signal to obtain a narrowband spectral parameter; spectrum estimating means for outputting a wideband spectral parameter using the narrowband spectral parameter as it is as a wideband spectral parameter; and the output wideband spectral parameter Synthesis means for generating a wideband audio signal from

  Alternatively, an analysis means for analyzing a narrowband speech signal to obtain a narrowband spectrum parameter, and converting the narrowband spectrum parameter into another region as necessary, performing transformation, and inversely transforming into the spectrum parameter region, thereby obtaining a wideband spectrum. Spectrum estimation means for outputting parameters and synthesis means for generating a wideband speech signal from the outputted wideband spectrum parameters are provided.

  Another wideband speech restoration apparatus includes a spectrum decoding unit that estimates a wideband spectral parameter from a narrowband speech code, and a synthesis unit that generates a wideband speech signal from the estimated wideband spectral parameter.

  Alternatively, spectrum decoding means for estimating a wideband spectrum parameter using a narrowband spectral code separated from a narrowband speech code, and a wideband for estimating a wideband sound source signal using a narrowband excitation code separated from the narrowband speech code A sound source decoding unit and a synthesizing unit for generating a wide band speech signal from the estimated wide band spectrum parameter and the wide band sound source signal are provided.

  Still further, as the wideband excitation decoding means, zero padding means for inserting a predetermined zero value into each sample interval of the narrowband excitation signal restored from the narrowband excitation code is used.

  Alternatively, the wideband excitation decoding means includes a wideband adaptive excitation decoding means for estimating a wideband adaptive excitation signal using a narrowband adaptive excitation code separated from the input narrowband speech code, and a narrowband separated from the input narrowband speech code. A wideband drive excitation decoding means for estimating a wideband drive excitation signal using the drive excitation code and an addition means for generating a wideband excitation signal from these estimated wideband adaptive excitation signal and wideband drive excitation signal.

  Alternatively, the wideband excitation decoding means includes: a wideband long period prediction code decoding means for estimating a wideband long period prediction code using a narrowband long period prediction code separated from an input narrowband speech code; and an input narrowband speech code Wideband long period prediction residual decoding means for estimating a wideband long period prediction residual signal using the separated narrowband long period prediction residual code, and these estimated wideband long period prediction code and wideband long period prediction residual signal And adding means for generating a broadband sound source signal.

  Alternatively, a narrowband amplitude information decoding means for estimating narrowband amplitude information using a narrowband excitation code separated from a narrowband speech code, and a narrowband spectrum code and narrowband amplitude information separated from the narrowband speech code Spectral decoding means for estimating at least wideband spectral parameters or wideband amplitude information using, and synthesizing means for generating wideband speech signals from the estimated wideband spectral parameters and wideband amplitude information or wideband sound source signals as necessary .

  Alternatively, a wideband speech decoding unit that estimates a wideband speech signal using a narrowband speech code and a postfilter unit that performs postfiltering on the decoded and estimated wideband speech signal are provided.

  The above-described wideband speech restoration apparatus synthesizes a wideband speech signal from the wideband spectrum parameter estimated using the narrowband spectrum parameter and the wideband sound source signal estimated using the narrowband sound source signal.

  Further, a wideband sound source signal is generated by inserting a predetermined number of zeros between samples of the narrowband sound source signal, and a wideband sound signal is synthesized using this and the estimated wideband spectrum.

  When estimating the wideband excitation signal, the narrowband adaptive excitation code and the narrowband driving excitation signal are calculated by analyzing the input narrowband excitation signal, and the wideband adaptive excitation signal estimated using this narrowband adaptive excitation code is calculated. And a wide-band driving sound source signal estimated using a narrow-band driving sound source were added to obtain a wide-band sound source signal. A broadband audio signal is synthesized using this and the estimated broadband spectrum.

  As another method for estimating a wideband sound source signal, an input narrowband sound source signal is analyzed to calculate a narrowband long period prediction code and a narrowband long period residual signal, and estimation is performed using the narrowband long period prediction code. The wideband long-period prediction signal and the wideband long-period residual signal estimated using the narrowband long-period residual signal were used as a broadband sound source signal. A broadband audio signal is synthesized using this and the estimated broadband spectrum.

  Another wideband speech restoration apparatus analyzes a narrowband speech signal to calculate a narrowband spectrum parameter, narrowband amplitude information, and a narrowband sound source signal, and uses the narrowband spectrum parameter and the narrowband amplitude information to Either or both of the parameters and the broadband amplitude information are estimated. Thereafter, a wideband audio signal is synthesized with these signals and the wideband sound source signal estimated from the narrowband sound source signal.

  In addition, other wideband audio restoration devices perform post filtering on a wideband audio signal estimated using a narrowband audio signal, and mainly process high frequency characteristics.

  In addition, another wideband speech restoration apparatus synthesizes a wideband speech signal using the narrowband spectrum parameter as a wideband spectrum parameter by extending the characteristics of the narrowband spectrum parameter.

  In addition, other wideband speech restoration apparatuses use up to a specific order of narrowband spectral parameters, and inversely convert them into corresponding spectral parameters to obtain wideband spectral parameters, which are used to synthesize wideband speech signals. .

  In addition, another wideband speech restoration apparatus generates a narrowband synthesized sound and estimates a wideband speech signal using a narrowband speech code, and converts the narrowband synthesized sound into a signal obtained by up-sampling the narrowband synthesized sound or the narrowband synthesized sound. A wideband audio signal is synthesized by adding signals obtained by extracting mainly high frequency band components other than the narrowband synthesized sound of the audio signal.

  Another wideband speech restoration apparatus synthesizes a wideband speech signal using a wideband spectrum parameter estimated using a narrowband spectral code and a wideband sound source signal estimated using a narrowband excitation code.

  Further, the wideband excitation signal is generated by inserting a predetermined number of zero values between each sample of the narrowband excitation decoded by the narrowband excitation code by the broadband excitation decoding means. A broadband audio signal is synthesized using the spectrum.

  Further, the wideband excitation signal is generated by adding the wideband adaptive excitation signal estimated using the narrowband adaptive excitation code and the wideband driving excitation signal estimated from the narrowband driving excitation signal by other broadband excitation decoding means. A broadband audio signal is synthesized using this and the estimated broadband spectrum.

  Further, the wideband excitation signal is generated from the wideband long period prediction code estimated using the narrowband excitation code and the wideband long period residual signal estimated from the narrowband long period prediction residual signal by other wideband excitation decoding means. Synthesized. A broadband audio signal is synthesized using this and the estimated broadband spectrum.

  In another wideband speech restoration apparatus, either or both of the wideband spectrum parameter and the wideband amplitude information are estimated using the narrowband spectrum code and the narrowband amplitude information. Thereafter, a wideband audio signal is synthesized with the information and the wideband sound source signal estimated from the narrowband sound source signal.

  In another wideband speech restoration apparatus, post-filtering is performed on a wideband speech signal estimated using a narrowband speech code, and mainly high frequency characteristics are processed.

  As described above, the wideband sound source signal is estimated using the narrowband sound source signal, and the wideband sound signal is synthesized using this, so that the characteristics of the narrowband sound source signal are given to the wideband sound source signal well. This is advantageous in that it can restore wideband speech with a stable and natural sound quality with less dependence on the speaker.

  Also, as the wideband sound source estimation means, zero padding means for inserting a predetermined number of zeros between each sample of the narrowband sound source signal is used, so there is no need for voiced / unvoiced determination and pitch extraction, and voiced / unvoiced determination errors and pitch extraction It is possible to estimate a good wide-band sound source that is not affected by errors, and to recover a wide-band sound having a stable and natural sound quality.

  In addition, as the broadband excitation estimation means, the broadband adaptive excitation signal and the broadband driving excitation signal are estimated using the narrowband adaptive excitation code and the narrowband driving excitation signal, and the broadband excitation signal is generated therefrom. Features relating to the strength and fluctuation of pitch periodicity possessed by the band sound source signal are reflected in the wide band sound source signal, and there is an effect that it is possible to restore wide sound with good sound quality without pulse-like sound.

  In addition, since the fundamental frequency and the frequency of its harmonic components are correctly aligned at integer multiples, the connection between the narrowband component and the restored wideband component in the wideband audio signal is good, and the practical nature of pitch periodicity can also be restored. This has the effect of restoring high-quality wideband audio.

  Also, as a broadband sound source estimation means, a wideband long period prediction code and a wideband long period residual signal are estimated using a narrowband long period prediction code and a narrowband long period residual signal, and the wideband sound source signal is estimated using these. Therefore, the characteristics of the pitch periodicity of narrow-band sound source signals and the characteristics of fluctuations are well reflected in the wide-band sound source signal, and there is no pulsing sound to restore wide-band sound with good sound quality. There is an effect that can.

  Furthermore, since the fundamental frequency and the frequency of its harmonic components are correctly aligned at integer multiples, the narrow-band component and the restored broadband component in the finally restored broadband audio signal are well connected, and the actual pitch periodicity Characteristics can also be incorporated, and there is an effect that high-quality wideband voice can be restored.

  Also, since one or both of the broadband spectrum parameter and the broadband amplitude information is estimated using the narrowband spectrum parameter and the narrowband amplitude information, the narrowband amplitude information is reflected in the estimation of the broadband spectrum parameter, A good spectrum can be estimated more stably, and there is an effect that wideband speech having a more correct amplitude can be restored.

  Furthermore, since post-filtering is performed on the wideband audio signal estimated using the narrowband audio signal, when the sound quality of the restored wideband audio signal is insufficient, the enhancement of pitch periodicity and the extreme of the spectral envelope are performed. This has the effect of improving sound quality such as emphasis.

  Furthermore, since the narrow band spectrum parameter is expanded and used as the wide band spectrum parameter to synthesize the wide band speech signal, there is an effect that the rough wide band spectrum can be restored very easily. In addition, there is no need for a memory for storing the code book, and the amount of calculation is reduced.

  Furthermore, since a wideband spectrum parameter is obtained by inversely converting it to a spectral parameter using up to a predetermined order of the narrowband spectral parameter, there is an effect that a rough broadband spectrum can be restored very easily. In addition, there is no need for a memory for storing the code book, and the amount of calculation is reduced.

  Further, according to the present invention, a narrowband synthesized sound is generated and a wideband speech signal is estimated using a narrowband speech code, and the narrowband synthesized signal is narrowed to a signal obtained by upsampling the narrowband synthesized sound or the narrowband synthesized sound. Since the components of the band other than the band synthesized sound are extracted and added, wideband speech can be restored from the encoded narrowband speech, and the decoded narrowband speech is not re-analyzed, so restoration is possible with a small amount of processing. There is an effect that can be done.

  Alternatively, since the wideband speech signal is synthesized using the wideband spectrum parameter estimated using the narrowband spectral code and the wideband sound source signal estimated using the narrowband excitation code, the decoded narrowband speech is reproduced again. There is no need for analysis, and there is an effect that restoration can be performed with a small amount of processing. Further, since distortion due to interpolation at the time of synthesis or windowing at the time of analysis is not superimposed, there is an effect that a broadband voice with better quality can be restored.

  Also, as the wideband sound source decoding means, zero padding means for inserting a predetermined number of zeros between each sample of the narrowband sound source decoded using the narrowband sound source code is used, so a sound source having an intermediate property between voiced and unvoiced Can be restored well, and there is an effect that it is possible to restore wideband sound of stable and natural sound quality.

  Also, as the wideband excitation decoding means, the wideband adaptive excitation signal estimated using the narrowband excitation code and the wideband driving excitation signal are estimated and added to obtain the wideband excitation signal. The wideband sound source signal is directly generated without performing the process, and there is an effect that it can be restored with a small amount of processing.

  In addition, since the characteristics relating to the strength and fluctuation of pitch periodicity included in the narrowband excitation code are well reflected in the broadband excitation signal, it is possible to restore wideband speech with good sound quality.

  Also, as the wideband excitation decoding means, the wideband long-period prediction code and the wideband long-period residual signal estimated using the narrowband excitation code are estimated, and the wideband excitation signal is synthesized using these. The wideband sound source signal is directly generated without decoding the narrowband sound source signal, and there is an effect that it can be restored with a small amount of processing.

  In addition, since the characteristics relating to the strength and fluctuation of pitch periodicity included in the narrowband excitation code are well reflected in the broadband excitation signal, it is possible to restore wideband speech with good sound quality.

  In addition, narrowband spectral information and narrowband amplitude information are used to estimate wideband spectral parameters and / or wideband amplitude information, so narrowband amplitude information is reflected in the estimation of wideband spectral parameters. Therefore, a good spectrum can be estimated more stably, and the difference in the narrowband spectrum code can be reflected in the estimation of the wideband amplitude information, so that it is possible to restore wideband speech having a more correct amplitude.

  Furthermore, since the post-filtering is performed on the wideband speech signal estimated using the narrowband speech code, when the postfilter process is applied to the narrowband synthesized sound, the narrowband portion and the restored band There is an effect of improving continuity. Further, when the sound quality of the restored wideband audio signal is insufficient, there is an effect that the sound quality can be improved, such as emphasizing pitch periodicity and emphasizing the pole of the spectrum envelope.

It is a block diagram of the wideband audio | voice restoration apparatus of Example 1 of this invention. It is explanatory drawing explaining the process of the zero padding means in Example 1 of this invention. It is a block diagram of the wideband sound source estimation means in the wideband audio | voice restoration apparatus of Example 2 of this invention. It is explanatory drawing explaining an example of the adaptive sound source signal in Example 2 of this invention. It is a block diagram of the wideband drive sound source estimation means in the wideband audio | voice restoration apparatus of Example 3 of this invention. It is a block diagram of the wideband drive sound source estimation means in the wideband audio | voice restoration apparatus of Example 4 of this invention. It is a block diagram of the wideband sound source estimation means in the wideband audio | voice restoration apparatus of Example 5 of this invention. It is a block diagram of the wideband drive sound source estimation means in the wideband audio | voice restoration apparatus of Example 6 of this invention. It is a block diagram of the wideband drive sound source estimation means in the wideband audio | voice restoration apparatus of Example 7 of this invention. It is a block diagram of the wideband sound source estimation means in the wideband audio | voice restoration apparatus of Example 8 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 9 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 10 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 11 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 12 of this invention. It is explanatory drawing explaining the relationship between the outline of a narrowband spectrum and a wideband spectrum in Example 13 of this invention. It is explanatory drawing explaining the relationship of the rough shape of the narrowband spectrum and wideband spectrum in Example 14 of this invention. It is a block diagram of the wideband spectrum estimation means in the wideband audio | voice restoration apparatus of Example 15 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 17 of this invention. It is a block diagram of the wideband sound-source decoding means in the wideband audio | voice restoration apparatus of Example 18 of this invention. It is a block diagram of the wideband sound-source decoding means in the wideband audio | voice restoration apparatus of Example 19 of this invention. It is a block diagram of the wideband sound-source decoding means in the wideband audio | voice restoration apparatus of Example 20 of this invention. It is a block diagram of the wideband audio | voice restoration apparatus of Example 22 of this invention.

Explanation of symbols

  1 narrowband speech signal, 2 analysis means, 3 spectrum analysis means, 4 narrowband spectrum parameters, 5 inverse filter, 6 narrowband excitation signal, 7 wideband spectrum estimation means, 8 vector quantization means, 9 narrowband spectrum codebook, DESCRIPTION OF SYMBOLS 10 Spectral code, 11 Inverse quantization means, 12 Wideband spectral codebook, 13 Wideband spectrum parameter, 14 Wideband sound source estimation means, 15 Zero padding means, 16 Wideband sound source signal, 17 Synthesis filter, 18 Bandpass filter, 19 Upsampling means, 20 Wideband speech signal, 21 Sound source analysis means, 22 Narrowband adaptive codebook, 23 Distortion minimization means, 24 Narrowband drive sound source signal, 25 Narrowband adaptive lag length, 26 Narrowband adaptive gain, 27 Wideband drive sound source estimation means, 28 Zero padding means, 29 Broadband drive sound source signal, 30 Band adaptive excitation estimation means, 31 Wideband adaptive excitation codebook, 32 Wideband adaptive excitation signal, 33 Wideband adaptive lag length, 34 Wideband adaptive gain, 35 Power calculation means, 36 Noise generation means, 37 Narrowband long period prediction analysis means, 38 Narrowband long period delay, 39 Narrowband long period prediction coefficient, 40 Long period inverse filter, 41 Narrowband long period prediction residual signal, 42 Wideband long period prediction residual estimation means, 43 Zero padding means, 44 Wideband long period prediction Parameter estimation means, 45 Wideband long period delay, 46 Wideband long period prediction coefficient, 47 Long period synthesis filter, 48 Wideband long period prediction residual signal, 49 Upsampling means, 50 Zeroing means, 51 Narrowband power calculation means, 52 Narrowband excitation power, 53 Narrowband power-included spectrum code, 54 Sound source normalization means, 55 Narrowband positive Normalized excitation signal, 56 Wideband normalized excitation signal, 57 Wideband power codebook, 58 Wideband excitation power, 59 Wideband excitation power estimation means, 60 Wideband power-included spectrum codebook, 61 Post filter means, 62 Spectral parameter conversion means, 63 Order reduction means, 64 spectral parameter inverse transformation means, 101 narrowband speech code, 102 separation means, 103 narrowband spectral code, 104 narrowband excitation code, 105 wideband spectrum decoding means, 106 wideband excitation decoding means, 107 narrowband spectrum decoding 108 narrowband excitation decoding means 109 narrowband speech decoding means 110 narrowband pitch code 111 narrowband power code 112 wideband pitch decoding means 113 wideband pitch period 114 wideband power decoding means 115 wideband Power decoding means, 116 excitation generating means, 117 narrowband adaptive excitation code, 118 narrowband driving excitation code, 119 wideband adaptive excitation decoding means, 120 wideband driving excitation decoding means, 121 narrowband adaptive excitation decoding means, 122 narrowband driving excitation Decoding means, 123 narrowband long period prediction code, 124 wideband long period prediction parameter decoding means, 125 narrowband long period prediction parameter decoding means, 126 narrowband long period prediction residual code, 127 wideband long period prediction residual decoding means, 128 narrowband long period prediction residual decoding means, 129 narrowband power decoding means, 130 wideband normalized excitation decoding means.

Claims (2)

A narrowband spectral decoding step for decoding narrowband spectral parameters from the narrowband speech code;
A narrowband excitation decoding step for decoding a narrowband excitation signal from the narrowband speech code;
A broadband excitation decoding step for generating a broadband excitation signal having the same power as the power of the narrowband excitation signal based on white noise;
A spectrum decoding step of outputting a broadband spectral parameter representing a spectral envelope reflecting a shape of a high frequency band of the spectral envelope represented by the narrow band spectral parameter in a high frequency shape;
A wideband speech restoration method comprising: a synthesis step of generating a wideband speech signal based on the broadband spectral parameter and the broadband sound source signal. Narrowband spectral decoding means for decoding narrowband spectral parameters from the narrowband speech code;
Narrowband excitation decoding means for decoding a narrowband excitation signal from the narrowband speech code;
Wideband sound source decoding means for generating a wideband sound source signal having the same power as the power of the narrowband sound source signal based on white noise;
Spectrum decoding means for outputting a broadband spectral parameter representing a spectral envelope reflecting a shape of a high frequency band of the spectral envelope represented by the narrow band spectral parameter in a high frequency shape;
A wideband speech restoration apparatus comprising: synthesis means for generating a broadband speech signal based on the broadband spectral parameter and the broadband sound source signal.

Images