JPH09101798A - Method and device for expanding voice band - Google Patents

Abstract

PROBLEM TO BE SOLVED: To synthesize a voice of high tone quality by converting a band-limited input voice into a broadband voice having a wide frequency band including the frequency band in the input voice. SOLUTION: This device is equipped with a voice analyzer 101 which separates the band-limited input voice into a narrow-band residue signal and a narrow-band spectrum envelope by taking a voice analysis, a residue band widening unit 102 which generates a broadband residue signal from the narrow- band residue signal, an envelope band widening unit 103 which estimates a broadband spectrum envelope from the narrow-band spectrum envelope, a voice synthesizer 104 which synthesizes a broadband synthesized voice from the broadband residue signal and broadband spectrum envelope, a filter 105 which extracts out-band components other than the frequency band that the input voice has from the broadband synthesized voice, and a voice superposing unit 106 which superposes the waveforms of the out-band components and input voice on the time base to synthesize a broadband voice having a frequency band including the frequency band that the input voice has.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention converts an input voice band-limited to a certain frequency band into a wide band voice having a wide frequency band including the frequency band of the input voice, thereby producing a high-quality voice. TECHNICAL FIELD The present invention relates to a voice band expanding method and a voice band expanding device.

[0002]

2. Description of the Related Art Most of analog telephone communication is N
Although it is conducted through the public line network under the jurisdiction of TT, the band is limited to 300 Hz to 3.4 kHz due to physical restrictions of the line, and the low frequency band below 300 Hz and 3.4
Sound quality deterioration occurs due to the loss of the high frequency band above kHz. Further, in digital voice communication such as that of a mobile phone, there is a principle limitation that the voice band is limited due to the limitation of the bit rate.

Therefore, in recent years, there has been a strong demand for a technique for improving the quality of telephone voice while keeping the line as it is, and research on this problem has recently been actively conducted. For example,
Yoshida, Abe: "Reconstruction method of wideband speech from narrowband speech by codebook mapping", Proceedings of ASJ, 1-
8-18, pp.179-180, (1993.3). This method is based on the correspondence between the narrowband speech and the wideband speech codebook, and indirectly by extracting the wideband code for the telephone speech code obtained by vector quantization (VQ) from the wideband codebook. Wideband speech is obtained by obtaining a wideband spectrum and synthesizing the speech using the pitch as a sound source.

[0004]

However, in the conventional method as described above, since the band is expanded by mapping the codebook, the synthesized sound is greatly deteriorated and the processing amount is relatively large.

The present invention solves the above-mentioned problems.
An analog telephone or a band that is band-limited by a communication line is created by converting an input voice band-limited to a certain frequency band into a wide-band voice having a wide frequency band that covers the frequency band of the input voice and synthesizing it. It is an object of the present invention to provide a voice band expanding method and a voice band expanding device that can expand the limited band of a mobile phone and improve the call quality.

[0006]

In order to achieve the above object, the present invention provides an input voice band-limited to a specific frequency band to a wide band voice having a wide frequency band including the frequency band of the input voice. A speech band expanding method for converting, wherein the input speech is linearly predictively analyzed for each constant frame to be separated into a narrow band residual signal and a narrow band spectrum envelope, and the narrow band residual signal is subjected to a wide band residual by nonlinear processing. Generating a difference signal, estimating a wideband spectrum envelope from the narrowband spectrum envelope, using a linear prediction synthesis method from the wideband residual signal and the wideband spectrum envelope,
It is characterized in that wideband speech having a wide frequency band including the frequency band of the input speech is synthesized.

The present invention is also a voice band expansion device for converting an input voice band-limited to a specific frequency band into a wide band voice having a wide frequency band including the frequency band of the input voice, A speech analyzer that performs linear prediction analysis of the input speech for each fixed frame to separate it into a narrowband residual signal and a narrowband spectrum envelope, and a sampling frequency by inserting zero for each sample of the narrowband residual signal. And a residual bandwidth expander for creating a band-expanded residual signal in which only the bandwidth of the narrow-band residual signal is expanded by further performing aliasing filtering, and the band-expanded residual signal is filtered. A pitch filter for extracting a pitch-containing residual signal obtained by cutting out only the frequency band in which the fundamental frequency component of the voice exists from the band expansion residual signal, Switch-containing residual signal to generate a wideband residual signal, a wideband residual band signal, an envelope wideband unit that estimates a wideband spectral envelope from the narrowband spectral envelope, the wideband residual signal and the wideband spectral envelope. And a speech synthesizer for synthesizing wide-band speech having a wide frequency band including the frequency band of the input speech by using the linear predictive synthesis method.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention has the above-described structure,
In order to convert an input voice band-limited to a certain frequency band into a wideband signal having a wide frequency band including the frequency band of the input voice, an input voice band-limited and
Since a wideband synthesized voice having a wider frequency band than the input voice created using the input voice is created, and the final wideband voice is obtained by superimposing the input voice and the wideband synthesized voice by the filtering process. With a simple configuration, it is possible to widen the band of high-performance band-limited voice.

[0009] The sophistication of communication terminals such as mobile phones is progressing, and at that time, discussion of call quality is being actively conducted.
The telephone is an important communication means for many people from its invention to the present day, and improving its quality is a very important research subject. SUMMARY OF THE INVENTION The present invention is a method and apparatus for restoring analog telephone voice band-limited to 0.3 to 3.4 kHz due to line characteristics or digital voice band-limited due to bit rate limitation to the original wideband signal. Is provided.

A first embodiment of the present invention for expanding a band-limited voice into a wide band will be described below.

FIG. 1 is a block diagram showing the overall construction of the first embodiment of the present invention. In FIG. 1, reference numeral 101 is a voice analysis of band-limited input voice for each frame (however, a frame is a unit time in which a voice signal is divided by a predetermined period) to analyze a narrow band residual signal and a narrow band spectrum envelope. A voice analyzer that separates, for example, LPC for each frame
This is a part for calculating a spectrum envelope and a residual signal obtained by analysis (linear prediction analysis). Here, as the spectral envelope obtained by the LPC analysis, for example, LPC coefficient, PARCOR coefficient, reflection coefficient, LSP coefficient, LPC
A cepstrum coefficient, an LPC mel cepstrum coefficient, etc. can be considered. Since these are all feature quantities expressing features on the spectrum of speech, any coefficient may be used. Further, the residual signal is the remaining information obtained by removing the information of the spectral envelope from the input voice, and can be said to well represent the pitch structure in the voice. Even if a pitch, a multi-pulse train, or an excitation codebook is used instead of the residual signal, these all correspond to the characteristic representation of the residual signal obtained after the LPC analysis of the speech, so any information is used. It doesn't matter. Where the linear prediction coefficient, PA
RCOR coefficient, reflection coefficient, LPC cepstrum coefficient, L
Regarding the feature quantities such as PC Mel cepstrum coefficient, it is described in detail in "Audio Digital Signal Processing (Upper) (Lower)", Corona Publishing Co., co-authored by LR Rabiner and RWSchafer, and LSP coefficient. Regarding, for example, FKSoong, BHJuang: "Line Spectrum Pa
ir (LSP) and Speech Data Compression ", Proc.ICASSP, 8
4, pp.1.10.1-1.10.4. Pitches and multi-pulse trains are described in detail in, for example, Furui: "Acoustic and Speech Engineering", Modern Science Co., Ltd., and regarding excitation codebooks, for example, Ono: "Recent progress in speech coding technology". , Journal of Acoustical Society of Japan, Vol. 48, No. 1, pp.52-59,
(1992).

On the other hand, even if a cepstrum analysis, a PSE analysis, a wavelet transform or the like is used as another method of speech analysis, it is still a method of separating / extracting the characteristic of the speech on the frequency axis. It doesn't matter what you do. For example, in the case of cepstrum analysis or PSE analysis, it can be realized by using a cepstrum coefficient extracted by a lifter as a spectrum envelope and using the rest as a residual signal. These analytical methods are already known, and for example, for cepstrum analysis, LRRa
Biner and RWSchafer's co-author, Kuki Suzuki's translation, "Digital signal processing of voice (top) (bottom)", Corona Publishing Co., Ltd.
Nakajima, Suzuki: “Power spectrum envelope (PSE) speech analysis and synthesis system”, Journal of Acoustical Society of Japan, Volume 44, No. 11, pp.824-
832, (1988), regarding wavelet transform, is described in Kawahara: “Application of Wavelet Analysis to Auditory Research”, Journal of Acoustical Society of Japan, Vol. 47, No. 6, pp.424-429, (1991). There is. In the present embodiment, hereinafter, LPC analysis is used as the voice analysis method and PAR is used as the spectrum envelope.
The COR coefficient is used as the residual signal.

Next, 102 is a residual band widening device for nonlinearly distorting the narrow band residual signal separated by the speech analyzer 101 to obtain a wide band residual signal. As a method of non-linearly distorting, first, a sampling frequency is doubled by inserting a zero for each sample of the narrow band residual signal, and further aliasing filtering is performed to obtain a sampled expanded residual signal without aliasing distortion. create. Of the sampling-enlarged residual signals, only those signals whose absolute values are below a certain value are changed to 0, or only those signals whose signal values are below 0 are changed to 0, or , Various methods such as a method of generating a wideband residual signal by inverting the sign of the value of a signal having a value of 0 or less are all possible. It can be used because it is a method of generating a signal. Further, if the absolute value of the sampling expanded residual signal has a value greater than a certain value and the sign is positive, the signal is changed to that certain value, and the absolute value of the signal has a value greater than a certain value, and If the sign is negative, the wideband residual signal can also be generated by changing the sign of a constant value to an inverted value. Furthermore, although the sound quality is slightly inferior, it is possible to artificially generate a wideband residual signal due to aliasing distortion by simply inserting a zero for each sample in the narrowband residual signal. It is possible. There are various methods that can be used to distort nonlinearly as described above, and the respective effects are greatly different.

Next, 103 is a narrow-band spectrum of the input voice separated by the voice analyzer 101 by a mapping function obtained by using a narrow-band spectrum envelope and a wide-band spectrum envelope extracted from a large amount of learning data in advance. An envelope broadening device that converts an envelope into a wideband spectral envelope. As the mapping function to be obtained, a linear mapping is used in the present embodiment, but a quadratic conversion other than the linear mapping or a non-linear conversion such as a neural network may be used to directly convert the wide band spectrum from the narrow band spectrum. It does not matter because they are the same. Regarding the secondary conversion, for example, F.Class, A.Kaltenmeier, P.
Regel, and K.Trottler: "Fast speaker adaptation for
speech recognition systems ", Proc. IEEE ICASSP, pp.
133-136, (Apr.1990), and conversion by a neural network, for example, Iso, Aso, Yoshida, Watanabe: "Speaker adaptation by neural network", Proceedings of the Acoustical Society of Japan, 1- 6-16, (1989.3),
It is described in. As the above-mentioned learning data, for example, various uttered voices of one standard speaker may be used, or by using data of a plurality of speakers,
This is useful when creating a mapping function that is robust against changes in the speaker's utterance.

Next, reference numeral 104 is a wide band having a wide frequency band including the frequency band of the input signal from the wide band residual signal and the wide band spectrum envelope obtained by the residual wide band expander 102 and the envelope wide band expander 103, respectively. A voice synthesizer for synthesizing synthetic voice. As the speech synthesis method, it is possible to decide what to use as the wideband residual signal and which feature amount to use as the wideband spectrum envelope. For example, when a residual signal or a pitch is used as the wideband residual signal and a linear prediction coefficient is used as the wideband spectrum envelope, the linear prediction synthesis method may be used for synthesis.
Besides, a cepstrum synthesis method, a PSE synthesis method, or the like can be used. Furthermore, it can be said that it is still better if the wide band residual signal output from the residual wide band expander 102 is subjected to spectrum smoothing processing before being synthesized by the voice synthesizer 104. For example, as an example of smoothing processing, by performing low-order linear prediction analysis on a wideband residual signal, spectrum flattening can be realized, and it is possible to reduce extra spectral unevenness when distortion occurs. Become. Further, there is a phase equalization process as a process close to the smoothing process, and by applying this to the wideband residual signal, it is possible to further improve the sound quality.

Reference numeral 105 is a filter for extracting frequency components other than the frequency band of the input voice from the wideband synthesized voice, and 106 is waveform superposition of the extracted out-of-band component and the input voice on the time axis. And a wideband voice having a wide frequency band including the frequency band of the input voice is synthesized.

The first embodiment of the present invention will be described in detail below with reference to the block diagram of FIG.

First, when a voice is input to the voice analyzer 101 after passing through a telephone line, a band limiting filter, etc., the voice analyzer 101 causes M-th order PAR at a constant time interval i.
The COR coefficient P _i (M) is extracted. Here, the fixed time interval is, for example, 8 KH for the sampling frequency of narrow band speech.
When z (bandwidth 4 KHz), 160 points (20 m
s), and this time unit is called a frame. On the other hand, the sampling frequency is set to 16
As KHz (bandwidth 8 KHz), 320 points (20 m
s).

Next, the narrow band residual signal separated by the voice analyzer 101 is nonlinearly distorted, and a wide band residual signal is obtained by the residual band widening unit 102. First, the sampling frequency is doubled by inserting zeros every other sample of the narrowband residual signal (upsampling). Then, by creating a filter (aliasing filter) that extracts only the signal component of the original frequency band and inputting the up-sampled residual signal to this filter,
Create a sampled expanded residual signal without aliasing. By doing so, for example, assuming that the original residual signal is data of 160 points in one frame, the sampling expanded residual signal has 320 points of data. And for example
A wideband residual signal can be generated by changing to 0 only those signals of which the absolute value of the signal has a value equal to or less than a fixed value among the sampling expanded residual signals.

On the other hand, the envelope broadening device 103 converts the narrow band spectral envelope of the input voice separated by the voice analyzer 101 into a wide band spectral envelope. First, a large number of training speech data prepared in advance are filtered to create a learning narrowband speech. As the filter, for example, a model simulating the characteristics of a telephone line or a model simulating a low-pass filter used in digitization is used. By this processing, the input voice and the learning voice can be treated as common data obtained in the same environment. Further, the wideband speech for learning is created without filtering the large number of learning data described above. The above processing procedure is shown in FIG.

Next, the details of the envelope broadening device 103 are shown in FIG. First, the narrow-band data processor 301 is a part that extracts the narrow-band spectrum envelope in the same manner as the speech analyzer 101 described above, using the learning narrow-band speech obtained earlier. Similarly, the wideband data processor 302 extracts a wideband spectrum envelope from the learning wideband speech described above. Next, the mapping function estimator 303 estimates the relationship between the extracted narrowband spectrum envelope and wideband spectrum envelope as a mapping function. In this embodiment, a linear mapping is used as the mapping function, and spectrum conversion from narrowband spectrum information to wideband spectrum information is performed. Specifically, the mapping function
The estimation of {A} is performed by minimizing the squared error of the difference between the wideband spectral envelope z _i after conversion of the input spectral envelope x _i and the target wideband spectral envelope y _i . That is, it is obtained by minimizing the objective function given by (Equation 1) in all frames of all learning data.

[0022]

(Equation 1)

However, {A} is an M × M dimensional matrix, and y _i and z _i are M dimensional vectors.

Since the wideband spectrum envelope and the narrowband spectrum envelope used in this estimation are obtained from the same learning speech data, it is possible to make a perfect one-to-one correspondence for each frame. Next, in step 304, the narrow band spectral envelope x _i of the input speech of the i-th frame extracted by the speech analyzer 101 is widened by the mapping function {A}.
_It is an envelope broadening device that converts to _i . Specifically, the conversion is performed according to (Equation 2).

[0025]

(Equation 2)

Next, the speech synthesizer 104 synthesizes speech by the linear predictive synthesis method from the wideband residual signal and the wideband spectrum envelope obtained by the residual wideband 102 and the envelope wideband 103, respectively. As the speech synthesis method, it is possible to decide what to use as the wideband residual signal and which feature amount to use as the wideband spectrum envelope.
For example, when the residual signal or the pitch is used as the wideband residual signal and the linear prediction coefficient is used as the wideband spectrum envelope, the linear prediction synthesis method may be used. Besides, a cepstrum synthesis method, a PSE synthesis method, or the like can be similarly used.

Next, the filter 105 extracts frequency components other than the frequency band of the input voice from the wideband synthesized voice. As a filter, an FIR filter or II
It does not matter which filter such as the R filter is used because it has the same meaning as extracting the frequency component other than the frequency band of the input voice from the wideband synthesized voice.

Finally, in the voice superimposing unit 106, waveforms of the extracted low-frequency components, high-frequency components, and other out-of-band components and the input voice are superimposed on the time axis to include the frequency band of the input voice. Wideband speech having a wide frequency band is synthesized.

As described above, according to the configuration of this embodiment,
It is possible to provide a voice band expanding method and device capable of accurately converting a band-limited input voice into a wide band voice having a wide frequency band with a relatively simple configuration.

Next, a second embodiment of the present invention will be described. This embodiment basically has the same configuration as that of the first embodiment (FIG. 1), but is an example in which a part thereof, that is, the residual band broadening device 102 is changed to another configuration.

FIG. 4 is a block diagram showing in detail the residual band broadening device of this embodiment. First, the voice analyzer 101
The input speech band-limited by is subjected to speech analysis for each frame and separated into a narrow band residual signal and a narrow band spectrum envelope. A major difference of the present embodiment from the first embodiment, that is, the most characteristic feature of the present embodiment is that when the separated narrowband residual signal is nonlinearly distorted to obtain a wideband residual signal, Before the band residual signal is distorted nonlinearly, the pitch filter 401 performs a filtering process to cut out only the frequency band in which the fundamental frequency component of the voice exists from the residual signal.

Next, the distortion generator 402 generates non-linear distortion to create a wide band residual signal. By extracting only the fundamental frequency component (pitch-containing residual signal) in the narrowband residual signal and generating non-linear distortion in this way, unnecessary distortion generated from signal components other than pitch can be reduced. It is possible to create a high-quality wideband residual signal. Here, the pitch filter is a filter used to cut out only a frequency band in which a fundamental frequency component of voice exists, and a low-pass filter having a cutoff frequency near 800 Hz is an example. This cutoff frequency can be arbitrarily determined.

In the following, the envelope band broadening device 103 converts the narrow band spectral envelope of the input voice separated by the voice analyzer 101 into a wide band spectral envelope, and further from the obtained wide band residual signal and wide band spectral envelope. Wideband synthetic speech having a wide frequency band including the frequency band of the input signal is synthesized by speech synthesis, and frequency components other than the frequency band of the input speech are extracted from the wideband synthetic speech by a filter, and the input speech and the time axis are The portion where the waveform is superimposed on the above to obtain a wide band voice having a wide frequency band including the frequency band of the input voice is the same as in the first embodiment.

Next, a third embodiment of the present invention will be described. This embodiment has the same basic configuration as that of the first embodiment (FIG. 1), and a detailed description of common parts will be omitted. The main difference of the present embodiment from the first embodiment is that in the first embodiment, the filter 105 extracts the frequency components other than the frequency band of the input voice from the wideband synthesized voice by the filter, and the voice superimposing unit 106 While the waveform was superimposed on the input voice as it is on the time axis, in the present embodiment, the components extracted by the filter and the input voice are each multiplied by a certain ratio, and then the waveform is superimposed on the time axis. Where it is.

The third embodiment of the present invention will be described in detail below. First, the speech analyzer 101 speech-analyzes the input speech for each frame to separate the narrowband residual signal and the narrowband spectrum envelope, and the residual wideband averaging unit 102 converts the narrowband residual signal into the wideband residual signal. The part for generating is the same as in the first embodiment. In this embodiment, the envelope broadening device 103 is different from that of the first embodiment.

FIG. 5 is a block diagram showing in detail the portion of the envelope broadening device 103. Hereinafter, description will be given with reference to the block diagram of FIG. First, 501 is a narrowband data generator that creates a narrowband spectrum envelope extracted in advance from a large amount of learning speech data, and 502 is a wideband spectrum temporally associated with this narrowband spectrum envelope. It is a narrow band data generator that creates an envelope. Next, the narrow band codebook generator 50
The narrowband spectrum envelope is classified into several similar spectrum envelopes by 3, and each representative code is obtained. Then, the mapping function estimator 504 estimates a mapping function for deriving a wideband spectrum envelope from the narrowband spectrum envelope for each representative code. The estimation method is the same as in the first embodiment. Then, the code determiner 505 determines to which representative code the narrow band spectrum envelope of the actual input speech is close.
Then, the envelope broadening unit 506 converts it into a wideband spectrum envelope using a mapping function corresponding to the closest code.

For example, in order to actually perform vector quantization on the narrow band spectral envelope, first, the quantization distortion D _jk for the k-th code V _k (code number L) for the j-th frame narrow band spectral envelope x _j is It is calculated by (Equation 3).

[0038]

(Equation 3)

However, x _j and V _k are M-dimensional vectors (M
Dimensional feature amount). The code with the smallest distortion is the representative code for the j-th frame narrow band spectrum envelope.

As described above, the method of clustering the narrow-band spectrum envelope into groups of similar spectra and obtaining some representative codes representatively representing each group is a vector quantization method (Y. Linde, A. buzo and RM
Gray: "An algorithm for vector quantizer design", IE
EE Trans.Commun., COM-28,1, pp.84-95 (Jan.1980)), which makes it possible to efficiently express the characteristics of a large amount of data with a small amount of data. Further, there is no problem even if another method is used as the clustering method.

From the wide band residual signal and the wide band spectrum envelope obtained as described above, the voice synthesizer 104 synthesizes a wide band synthesized voice having a wide frequency band including the frequency band of the input voice. This part is the same as in the first embodiment.

Furthermore, in this embodiment, the filter 105 and the voice superimposing unit 106 of the first embodiment are different. FIG. 6 is a block diagram showing in detail the filter and the voice superimposing unit. Hereinafter, description will be given with reference to FIG.

First, 601 is a low-pass filter for extracting only low-frequency components of the frequency components other than the frequency band of the input voice from the wide-band synthesized voice by a low-pass filter, and 602 is the input voice from the wide-band synthesized voice. It is a high-pass filter that extracts only high-frequency components from the frequency components other than the frequency band that it has by a high-pass filter. Here, two types of filters, low band and high band, are considered as filters for extracting frequency components other than the frequency band of the input voice from the wideband synthesized voice, but three or more filters may be used without any problem.

Reference numeral 603 denotes a voice superimposing device which superimposes the low frequency component and the high frequency component and the input voice on the time axis.
The present embodiment has a function of superimposing the low frequency component and the high frequency component on the input voice at a predetermined ratio according to the content of the code determined to be the closest code by the code determination unit 503. At this time, for example, in the case where the chord is an unvoiced sound such as a fricative sound, the ratio of the low frequency range, the high frequency range, and the input voice is set to 0.5: 1: 1.5 so that the high frequency range is When the chord is a voiced voice such as a vowel, emphasize the low range by setting the ratio of low range, high range, and input voice to 1.5: 1: 0.5. By superimposing the waveform on the input voice, it is possible to accurately widen the band of the input voice according to the shape of the code, that is, the voice spectrum.

Next, a fourth embodiment of the present invention will be described. This embodiment has many parts in common with the third embodiment, and detailed description of the common parts will be omitted.
The difference of the present embodiment from the third embodiment is that in the third embodiment, when the voice superimposing unit 603 superimposes the low frequency component and the high frequency component on the input voice on the time axis, the code determining unit 503 According to the content of the code determined to be the closest code, the low-frequency component, the high-frequency component and the input voice are superposed at a ratio determined in advance. Is multiplied by the ratio according to the voicedness of the input voice, and then the waveform is superimposed on the time axis.

The fourth embodiment of the present invention will be described in detail below. First, the speech analyzer 101 speech-analyzes the input speech for each frame to separate the narrowband residual signal and the narrowband spectrum envelope, and the residual wideband averaging unit 102 converts the narrowband residual signal into the wideband residual signal. Then, the envelope band broadening unit 103 estimates the wide band spectrum envelope from the narrow band spectrum envelope. Further, the speech synthesizer 104 synthesizes wideband synthesized speech having a wide frequency band including the frequency bandwidth of the input speech. The parts up to this point are the same as in the third embodiment.

However, this embodiment is different in the part of the voice superimposing unit 106 of the third embodiment. FIG. 7 is a block diagram showing in detail the parts of the voice superimposing unit and the filter. Hereinafter, description will be given with reference to FIG. 7.

From the wide-band synthesized speech synthesized by the speech synthesizer 104, a low-pass filter 701 extracts only low-frequency components of frequency components other than the frequency band of the input speech, and a high-pass filter 702 extracts the input speech. Only the high frequency component is extracted from the frequency components other than the frequency band that it has. This part is similar to that of the third embodiment.

Next, in the present embodiment, first, the voiced utterance determiner 703 determines the voicedness of the input voice using the narrowband residual signal and the narrowband spectrum envelope extracted from the input voice. Further, the low-frequency component, the high-frequency component, and the input voice are superimposed by the voice superimposing unit 704 at a ratio determined in advance in accordance with the voiced ratio. At this time, as the ratio, for example, when voicedness indicates unvoiced sound such as fricative, the high frequency is emphasized by setting the ratio of low frequency, high frequency, and input voice to 0.5: 1: 1.5. If the voicedness indicates a voiced sound such as a vowel, the ratio of the low frequency range, the high frequency range, and the input voice is set to 1.5: 1: 0.5 so that the low frequency range is emphasized. By superimposing the waveform on the voice, it is possible to accurately widen the bandwidth of the input voice according to the voicedness, that is, the shape of the spectrum. As the voiced characteristic of the input voice, for example, the autocorrelation coefficient of the input voice is obtained, and the normalized autocorrelation coefficient divided by the 0th-order coefficient, that is, the power is obtained. It can be realized by determining that the pitch coefficient corresponds to the pitch and using this pitch coefficient as voiced.

As described above, according to the configuration of this embodiment,
It is possible to provide a voice band expansion device capable of accurately converting a band-limited input voice into a wide band voice having a wide frequency band.

As described above, according to the voice band expanding device of the embodiment of the present invention, the input voice whose band is limited to a specific frequency band is voice analyzed for every constant frame, and the narrow band residual signal and the narrow band spectrum are analyzed. Speech analyzer 101 separated into envelope
A residual band broadening device 102 for generating a wide band residual signal from the narrow band residual signal, an envelope wide band device 103 for estimating a wide band spectrum envelope from the narrow band spectrum envelope,
A speech synthesizer 104 for synthesizing a wideband synthesized speech from a wideband residual signal and a wideband spectrum envelope, a filter 105 for extracting out-of-band components other than the frequency band of the input speech from the wide-band synthesized speech, an out-of-band component and the input speech. And a voice superimposing unit 106 for synthesizing a wide-band voice having a frequency band including the frequency band of the input voice by superimposing the waveforms on the time axis. A magnifying device can be provided.

[0052]

As is apparent from the above embodiments, according to the present invention, the input voice band-limited to the specific frequency band is voice-analyzed for every constant frame, and the narrow band residual signal and the narrow band spectrum are obtained. Separated into an envelope, generate a wideband residual signal from the narrowband residual signal, estimate a wideband spectrum envelope from the narrowband spectrum envelope, a linear prediction synthesis method from the wideband residual signal and the wideband spectrum envelope. By using, to synthesize a wide-band synthesized voice having a wide frequency band including the frequency band of the input voice, the frequency component other than the frequency band of the input voice is extracted from the wide-band synthesized voice by a filter, the extracted band Waveform superposition of the external component and the input voice on the time axis to synthesize a wideband voice having a wide frequency band including the frequency band of the input voice. Since the configuration, a relatively simple configuration, the input speech is band limited, it can be converted accurately into a broadband signal having a wide frequency band so as to encompass the frequency band possessed by the input speech.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the overall configuration of a voice band expansion device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a processing procedure for generating learning band speech in the embodiment.

FIG. 3 is a block diagram showing a configuration of an envelope broadband device according to the first embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a residual band broadening device according to a second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of an envelope broadening device according to a third embodiment of the present invention.

FIG. 6 is a block diagram of a filter and a voice superimposing device according to a third embodiment of the present invention.

FIG. 7 is a block diagram of a filter and a voice superimposing device according to a fourth embodiment of the present invention.

[Explanation of symbols]

 101 Speech Analyzer 102 Residual Bandwidth Extender 103 Envelope Bandwidth Extender 104 Speech Synthesizer 105 Filter 106 Speech Superimposer

Claims (22)

[Claims] 1. A voice band expanding method for converting an input voice band-limited to a specific frequency band into a wide band voice having a wide frequency band including the frequency band of the input voice, wherein the input voice is constant. A narrow band residual signal and a narrow band spectrum envelope are separated by linear prediction analysis for each frame, a wide band residual signal is generated from the narrow band residual signal by nonlinear processing, and the narrow band spectrum envelope is converted to a wide band spectrum envelope. And a wideband speech having a wide frequency band including a frequency band of the input speech is synthesized by using a linear prediction synthesis method from the wideband residual signal and the wideband spectrum envelope. Expansion method. 2. A method of generating a wideband residual signal from a narrowband residual signal by non-linear processing, by inserting a zero for each sample in the narrowband residual signal to double the sampling frequency and wideband residual signal. The voice band expanding method according to claim 1, wherein a signal is generated. 3. As a method for generating a wideband residual signal from a narrowband residual signal by nonlinear processing, a sampling frequency is doubled by inserting a zero for each sample of the narrowband residual signal, and further aliasing is performed. A band-expanded residual signal in which only the bandwidth of the narrow-band residual signal is expanded by filtering is created, and the absolute value of the band-expanded residual signal has a positive positive value or more and the signal If the sign of is positive, the signal is changed to the constant value, and if the absolute value of the signal has a value equal to or greater than the constant value and the sign of the signal is negative, the sign of the constant value is inverted. The method for expanding a voice band according to claim 1, wherein the wide band residual signal is generated by changing the signal. 4. A method for generating a wideband residual signal from a narrowband residual signal by non-linear processing, by inserting a zero for each sample of the narrowband residual signal to double the sampling frequency and further aliasing. A band-expanded residual signal in which only the bandwidth of the narrowband residual signal is expanded by filtering is created, and only a signal whose absolute value of the band-expanded residual signal has a value equal to or less than a certain value is generated. The method of claim 1, wherein a wide band residual signal is generated by changing the value to 0. 5. A method for generating a wideband residual signal from a narrowband residual signal by non-linear processing, by inserting a zero for each sample of the narrowband residual signal to double the sampling frequency and further aliasing. A band expansion residual signal in which only the bandwidth of the narrow band residual signal is expanded by filtering is created, and only signals having a signal value of 0 or less in the band expansion residual signal are set to 0. The method for expanding a voice band according to claim 1, wherein the wide band residual signal is generated by changing the signal. 6. A method of generating a wideband residual signal from a narrowband residual signal by non-linear processing, in which a sampling frequency is doubled by inserting a zero for each sample of the narrowband residual signal, and further aliasing is performed. A band expansion residual signal in which only the bandwidth of the narrow band residual signal is expanded by filtering is created, and only a signal having a value of 0 or less in the band expansion residual signal 2. The voice band expanding method according to claim 1, wherein the wide band residual signal is generated by inverting the sign. 7. A method for estimating a wide band spectrum envelope having a wide frequency band including the frequency band of the input voice from a narrow band spectrum envelope obtained by voice analysis of the input voice band-limited to a specific frequency band. , A wide band having a wide frequency band including a narrow band spectrum envelope having the same frequency band as the input voice extracted in advance from a large amount of learning voice data and a frequency band of the input voice correlated in time A spectral envelope is used to estimate a mapping function for deriving the wideband spectral envelope from the narrowband spectral envelope for each of several similar spectral envelopes, and the mapping function is used to estimate the narrowband spectral envelope of the input speech. Wideband spectrum having a wide frequency band including the frequency band of the input voice based on The speech band expanding method according to claim 1, wherein the envelope is estimated. 8. A voice band expanding device for converting an input voice band-limited to a specific frequency band into a wide band voice having a wide frequency band including the frequency band of the input voice, wherein the input voice is constant. A speech analyzer that performs linear prediction analysis for each frame to separate it into a narrowband residual signal and a narrowband spectrum envelope, and doubles the sampling frequency by inserting zero for each sample of the narrowband residual signal, A residual bandwidth expander that creates a band-expanded residual signal by expanding only the bandwidth of the narrowband residual signal by further performing aliasing filtering, and the band-expanded residual signal by filtering the band-expanded residual signal. A pitch filter for extracting a pitch-containing residual signal obtained by cutting out only a frequency band in which a fundamental frequency component of voice exists from the band-expanded residual signal, and the pitch-containing residual signal. Signal to generate a wideband residual signal, an envelope wideband device that estimates a wideband spectrum envelope from the narrowband spectrum envelope, and a linear prediction combining method from the wideband residual signal and the wideband spectrum envelope. And a voice synthesizer for synthesizing wide-band voice having a wide frequency band including the frequency band of the input voice by using the voice band expanding device. 9. The residual widening device changes the signal to the constant value if the absolute value of the signal among the pitch-containing residual signals has a value greater than or equal to a constant value and the sign of the signal is positive, A wideband residual signal is generated by changing the sign of the constant value to an inverted value if the absolute value of the signal has a value equal to or greater than a fixed value and the sign of the signal is negative. 8. The voice band expansion device described in 8. 10. A residual wideband generator generates a wideband residual signal by changing to 0 only a signal having an absolute value of the signal below a fixed value among pitch-containing residual signals. The voice band expansion device according to claim 8. 11. The residual band widening device generates a wide band residual signal by changing only a signal having a value of 0 or less among pitch-containing residual signals to 0. The voice band expansion device according to claim 8. 12. A residual band widening device generates a wide band residual signal by inverting the sign of the value of only the signal having a value of 0 or less among the pitch-containing residual signals. The voice band expansion device according to claim 8. 13. An envelope broadening device temporally correlates the narrow band spectrum envelope having the same frequency band as the input voice extracted in advance from a large amount of learning voice data with the narrow band spectrum envelope. And estimating a mapping function for deriving the wideband spectral envelope from the narrowband spectral envelope for each of several similar spectral envelopes with a wideband spectral envelope having a wide frequency band including the frequency band of the input speech. Aside from that, the mapping function is used to convert the narrow-band spectrum envelope of the input voice into a wide-band spectrum envelope having a wide frequency band including the frequency band of the input voice. The voice band expansion device according to any one of 1. 14. A filter for extracting an out-of-band component of a frequency other than the frequency band of the input voice by a filter from the wide-band synthesized voice synthesized by the voice synthesizer, and the extracted out-of-band component and the input voice. 13. A voice superimposing device for synthesizing a wide-band voice having a wide frequency band including a frequency band of an input voice by superimposing a waveform on a time axis, and further comprising a voice superimposing device. A voice band expansion device according to claim 1. 15. A low-pass filter for extracting only a low-frequency component of a frequency component other than a frequency band of an input voice from a wide-band synthesized voice synthesized by a voice synthesizer by a low-pass filter, and the wide-band synthesized voice. From the high-pass filter that extracts only the high-frequency component of the frequency components other than the frequency band of the input voice by the high-pass filter, and the extracted low-pass component, the high-frequency component and the input voice are constant. 13. A voice superimposing device for synthesizing a wide-band voice having a wide frequency band including a frequency band of an input voice by superimposing a waveform on a time axis by a ratio. The voice band expansion device according to any one of the above. 16. An envelope broadening device, when estimating a wide band spectrum envelope from a narrow band spectrum envelope, has a narrow band spectrum envelope having the same frequency band as the input voice extracted from a large amount of learning voice data in advance. To create a codebook by classifying each of several similar spectral envelopes, and further obtain a representative code representative of each codebook, and a narrowband spectral envelope extracted from the learning voice data, and Using the wideband spectrum envelope having a wide frequency band including the frequency band of the input speech temporally associated with the narrowband spectrum envelope extracted from the learning voice data, the narrowband spectrum envelope to the wideband The mapping function for deriving the spectrum envelope is estimated for each representative code, and the narrow band spectrum envelope of the input speech is calculated. A wideband spectrum having a wide frequency band including the frequency band of the input voice is determined by determining which of the representative codes is closest, determining the most similar code as a similar code, and using the mapping function corresponding to the similar code. The speech synthesizer uses a linear predictive synthesis method from the wideband residual signal and the wideband spectrum envelope to synthesize a wideband synthesized speech having a wide frequency band including the frequency band of the input speech. Further, a low-pass filter for extracting only a low-frequency component of the frequency components other than the frequency band of the input voice from the wide-band synthesized voice by a low-pass filter, and the input voice from the wide-band synthesized voice. The high-pass filter that extracts only the high-frequency components of the frequency components other than the frequency band of the When the output low-frequency component, the high-frequency component and the input voice are waveform-superimposed on the time axis, the low-frequency component, the aforesaid ratio at a ratio determined in advance according to the content of the similar code. 9. A voice superimposing device for synthesizing a wide band voice having a wide frequency band including a frequency band of the input voice by superposing a high frequency component and the input voice. The voice band expansion device according to claim 12. 17. The voice superimposing device, if the similar code is a code representing an unvoiced sound such as a fricative sound, the low frequency band of the input voice is 0.5.
If the similar code is a code representing a voiced sound such as a vowel, the low range is set to 1.5 times the input voice and the high range is set to 1.5 times the input voice. 17. The voice band expanding device according to claim 16, wherein the frequency band is set to 0.5 times the input voice so as to emphasize the low frequency region and the waveform is superimposed on the input voice. 18. The envelope broadening device, when estimating the wide band spectrum envelope from the narrow band spectrum envelope, has a narrow band spectrum envelope having the same frequency band as the input voice extracted from a large amount of learning voice data in advance. To create a codebook by classifying each of several similar spectral envelopes, and further obtain a representative code representative of each codebook, and a narrowband spectral envelope extracted from the learning voice data, and Using the wideband spectrum envelope having a wide frequency band including the frequency band of the input speech temporally associated with the narrowband spectrum envelope extracted from the learning voice data, the narrowband spectrum envelope to the wideband The mapping function for deriving the spectrum envelope is estimated for each representative code, and the narrow band spectrum envelope of the input speech is calculated. A wideband spectrum having a wide frequency band including the frequency band of the input voice is determined by determining which of the representative codes is closest, determining the most similar code as a similar code, and using the mapping function corresponding to the similar code. The speech synthesizer uses a linear predictive synthesis method from the wideband residual signal and the wideband spectrum envelope to synthesize a wideband synthesized speech having a wide frequency band including the frequency band of the input speech. Further, a low-pass filter for extracting only a low-frequency component of the frequency components other than the frequency band of the input voice from the wide-band synthesized voice by a low-pass filter, and the input voice from the wide-band synthesized voice. The high-pass filter that extracts only the high-frequency components of the frequency components other than the frequency band of the When waveform-superimposing the output low-frequency component, the high-frequency component, and the input voice on the time axis, the voicedness of the input voice is obtained in advance from the narrowband residual signal, and At a ratio according to the low frequency component,
By superimposing the high frequency component and the input voice,
The voice band expanding device according to any one of claims 8 to 12, further comprising: a voice superimposing device that synthesizes a wide band voice having a wide frequency band including a frequency band of the input voice. 19. A voice superimposing device, if the input voice is a phoneme close to an unvoiced sound such as a fricative, sets the low frequency to 0.5 times the input voice and the high frequency to 1.5 times the input voice. When the high frequency range is emphasized and the input voice is close to a voiced sound such as a vowel, the low frequency range is 1.5 times the input voice and the high frequency range is 0.5 times the input voice.
19. The voice band expanding device according to claim 18, wherein the input voice is waveform-superimposed so as to emphasize the low frequency band by doubling the frequency. 20. A speech superimposing device obtains an autocorrelation coefficient of the input speech as a voiced characteristic of the input speech, and further obtains a zero-order coefficient, that is, a normalized autocorrelation coefficient divided by power,
It is characterized in that the maximum coefficient among the values other than the zero-order coefficient is determined to be the pitch coefficient corresponding to the pitch, and a waveform obtained by multiplying the pitch coefficient by an appropriate value is used to superimpose the waveform on the input voice. The voice band expansion device according to claim 17. 21. An FIR filter is used as the filter according to claim 8, claim 14 or claim 15,
The voice band expansion device according to any one of claims 16 and 18. 22. An IIR filter is used as the filter, claim 8, claim 14, claim 15,
The voice band expansion device according to any one of claims 16 and 18.