JP2005055778A - Equalizer of frequency characteristic of speech - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an equalizer of frequency characteristics of speech capable of stably correcting the change in the transmission characteristics of a sound recording system while taking human auditory characteristics into consideration. <P>SOLUTION: The equalizer 20 includes a source PSD forming section 34 which calculates the difference of the power spectral density (PSD) between source speech 30 and target speech 32, a target PSD forming section 36 and a dividing section 38, and a cepstrum filter section 40 which filters the source speech 30 by a filter parameter smoothed by using cepstrum transformations 70 and 72 and mel warping transformation 74 for the difference, and reverse transformation thereof 76. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech correction technique, and more particularly to a technique for correcting the sound quality of an input sound to a sound quality close to a target sound in a waveform segment connection type speech synthesis system or the like.

  With the development of computer technology and data communication technology, the interface between humans and machines has become important. For human beings, it is desirable to be able to communicate with machines in the same way as talking to people, and technology development for that purpose is underway.

  Cognitive techniques such as voice recognition and image recognition are mainly used for transmitting information from humans to machines. There are various methods for transmitting information from a machine to a human being, and among them, the opportunity to use speech synthesis technology is increasing. A typical application is a voice response system, a voice translation system, a computer game, and the like. Furthermore, with the development of robots and the like in recent years, it is expected that communication between humans and robots can be realized in the same way as communication between humans by combining voice recognition and image recognition with voice synthesis.

  In speech synthesis, how to synthesize natural speech is important. Recently, a speech unit is created by using a large speech corpus of several tens of hours to create speech units and select and connect appropriate speech units according to the input text data. Type speech synthesis has become mainstream. In this technique, it is important how to connect speech waveform segments naturally.

  As described above, in the current waveform segment speech synthesis system, a large-scale speech corpus is used to improve sound quality. In many cases, a single speaker's voice is recorded over a long period of time. In some cases, the recording may take months to years.

  In such a case, if the recording time is different, the characteristics of the recording system will change over time, and the sound quality may change if the recorded sound is reproduced. In the case of performing waveform connection, there is a problem in that synthesized speech becomes unnatural when speeches having different sound qualities are connected.

  One proposal has been made in Non-Patent Document 1 regarding a sound correction technique for solving such a problem. FIG. 4 shows a block diagram of the channel equalizer described in Non-Patent Document 1. Referring to FIG. 4, this apparatus 200 receives source sound 30 and generates sound having the same frequency characteristics as target sound 32 in the utterance content of source sound 30. In the present specification, the target sound 32 refers to a reference sound that has been recorded in advance. The source sound 30 is sound recorded at a different time from the target sound 32, and it is assumed that the frequency characteristic may be different from that of the target sound 32 due to the secular change of the characteristics of the recording system.

  The apparatus 200 includes a source PSD generation unit 34 for generating a power spectral density (PSD) of the source sound 30, a target PSD generation unit 36 for generating a PSD of the target sound 32, and a PSD of the source sound 30. A division unit 38 for calculating a difference (target PSD / source PSD) from the PSD of the target speech 32, and an IIR (Infinite Impulse Response) using a result of LPC (Linear Prediction Coefficient) analysis based on the output of the division unit 38 An LPC filter unit 210 for filtering the source audio 30 with a filter and outputting the equalized audio 212.

  The source PSD generation unit 34 and the target PSD generation unit 36 have the same configuration. The source PSD generation unit 34 performs a predetermined windowing process on each audio frame detected by the audio frame detection unit 50 for detecting each audio frame included in the data of the source audio 30 and the audio frame detection unit. A windowing processing unit 52, a power spectrum calculating unit 54 for calculating PSD of the sound frame by fast Fourier transform (FFT) from the sound frame data windowed by the windowing processing unit 52, and power spectrum calculation And a frame averaging unit 56 for calculating the average PSD of the frames of the source audio 30 for a predetermined period calculated by the unit 54.

  Similarly, the target PSD generation unit 36 includes an audio frame detection unit 60, a windowing processing unit 62, a power spectrum calculation unit 64, and a frame averaging unit 66.

  The LPC filter unit 210 includes an IFFT unit 220 for performing inverse FFT (IFFT) processing on the output of the division unit 38, an LPC conversion unit 222 for performing LPC conversion on the output of the IFFT unit 220, and an LPC conversion unit 222. And an IIR filter 224 for performing a filtering process on the source sound 30 to equalize the frequency characteristic of the source sound 30 to the frequency characteristic of the target sound 32.

  A frequency characteristic difference between the source sound 30 and the target sound 32 is calculated by the division unit 38, and LPC conversion is performed on the difference to set an IIR filter parameter. The channel equalizer 200 can equalize the frequency characteristic of the source sound 30 to be substantially equal to that of the target sound 32.

Yushi, Eric Chang, Hupen, Minchu, “Channel Equalization Based on Power Spectral Density for Large-Scale Speech Databases for Connected TTS Systems”, ICSLP2002 Proceedings, pp. 2369-2372, USA, 2002 (Yu Shi, Eric Chang, Hu Peng, and Min Chu, “Power Special Density Based Channel 7 Proc. )

  The effectiveness of the conventional equalization apparatus shown in FIG. 4 is shown.

  However, the conventional method has a difficult problem as to how to select the order in the LPC conversion. That is, if the order of the LPC conversion is reduced, there is a problem that the effect of the correction is almost lost, while if the order is increased, the sound quality is greatly deteriorated. Therefore, it is difficult to determine the LPC conversion order to an appropriate value.

  An apparatus for equalizing frequency characteristics of speech according to the present invention includes means for calculating a difference in power spectral density (PSD) between speech to be processed and reference speech, and the difference is represented by a cepstrum. And a predetermined filtering means for filtering the speech to be processed in which filter parameters are set using the difference represented by the cepstrum. .

  Preferably, the filtering means includes means for converting a difference cepstrum into a mel cepstrum, and MLSA (mel) connected so as to receive a voice to be processed as an input using the difference represented by the mel cepstrum as a filter parameter. -Logarithmic (special application) filter.

  Alternatively, the filtering means is a PSD difference smoothing means for performing a smoothing process on the PSD difference, and a sound to be processed in which a filter parameter is set by the PSD smoothed by the PSD difference smoothing means. And a FIR (Finite Impulse Response) filter connected to receive.

  The audio frequency characteristic equalizer may further include means for converting the PSD difference into a differential cepstrum having a predetermined first order. The PSD difference smoothing means performs a first melwarping for frequency axis warping from the first order cepstrum to a mel cepstrum having a second order smaller than the first order for the difference cepstrum. And second melwarping conversion means for performing inverse transformation of melwarping by the first melwarping means on the output of the first melwarping means and outputting a cepstrum of the inversely transformed difference, And a means for outputting a PSD difference smoothed by converting the cepstrum of the inversely converted difference output from the two melwarping conversion means into a spectrum and giving the difference as a filter parameter to the FIR filter.

  A difference in power spectral density between the speech to be processed and the speech to be processed is converted into a feature space having a frequency characteristic represented by a cepstrum. This is further converted into a mel scale and an MLSA filter is set. Alternatively, after converting to a cepstrum, Melwarping and its inverse transform are performed to obtain an inversely transformed cepstrum, which is further returned to the spectrum to smooth the PSD difference, and a filter is set with the PSD difference. . The filter set in this way has characteristics close to human auditory characteristics. Further, the characteristics of the filter set in this way are not sensitive to the parameter order, unlike the filter by LPC conversion. Even when the parameters are calculated so as to increase the accuracy of the filter, the sound quality does not deteriorate. Further, it is possible to equalize the frequency characteristic of the voice to be processed to the reference frequency characteristic with the same sound quality as that of the conventional channel equivalent apparatus.

[First Embodiment]
FIG. 1 shows a block diagram of a channel equivalent apparatus 20 according to the first embodiment of the present invention. In FIG. 1, the same components as those in FIG. 4 are denoted by the same reference numerals. Their functions are also the same. Therefore, detailed description thereof will not be repeated.

  The channel equivalent device 20 shown in FIG. 1 is different from the channel equivalent device 200 shown in FIG. 4 in that an FIR (Finite Impulse Response) set by a filter parameter obtained by smoothing the PSD difference is used instead of the LPC filter unit 210 shown in FIG. ) A cepstrum filter unit 40 that performs equalization using a filter is included.

  The cepstrum filter unit 40 takes a square root of the average difference between the PSDs of the source sound 30 and the target sound 32 output from the division unit 38 and calculates the logarithm thereof, and the output of the logarithm calculation unit 70. The IFFT processing unit 72 for calculating the cepstrum for the output of the division unit 38 by executing m-th order IFFT processing on the cepstrum, and the horizontal axis (frequency axis) of the cepstrum output from the IFFT processing unit 72 on the mel scale And a first melwarping unit 74 for converting to melwarping.

  The conversion in the first melwarping unit 74 is expressed as melwarping (m, n, a), where m is the order of the cepstrum before conversion, and n is the order after conversion. Here, m and n are selected so that n <m. a is a constant which is a parameter for frequency axis expansion and contraction, and is determined according to the sampling frequency.

  The cepstrum filter unit 40 further includes a second melwarping unit 76 for performing melwarping (n, m, -a) on the output of the first melwarping unit 74. The melwarping in the second melwarping section 76 and the melwarping in the first melwarping section 74 are in a mutually inverse relationship. That is, by executing these two processes in series, the frequency axis returns to the original linear axis. However, since the values of m and n are selected so that n <m, when these two processes are executed in series, the cepstrum of the high value portion on the frequency axis after the mel conversion is removed.

  The cepstrum filter unit 40 further includes an FFT processing unit 78 for performing FFT processing on the output of the second melwarping unit 76, an exponent conversion unit 80 for performing exponent conversion on the output of the FFT processing unit 78, An IFFT processing unit 82 for performing an IFFT process on the output of the exponent conversion unit 80 and outputting a filter parameter. The output of the IFFT processing unit 82 is obtained by smoothing the PSD of the output of the division unit 38 by melwarping by the first melwarping unit 74 and the second melwarping unit 76 and its inverse transformation.

  The cepstrum filter unit 40 is further set by a filter parameter output from the IFFT processing unit 82, and performs a filter process on the source sound 30, thereby correcting the frequency characteristic of the source sound 30 and substantially equal to the frequency characteristic of the target sound 32. An FIR 84 for outputting as sound 42 having the same frequency characteristic is included.

  The channel equivalent device 20 operates as follows. Referring to FIG. 2, source PSD 100 is obtained from waveform 90 of source audio 30 by audio frame detection unit 50, windowing processing unit 52, power spectrum calculation unit 54, and frame averaging unit 56 shown in FIG. 1. Similarly, the target PSD 102 is obtained by the audio frame detection unit 60, the windowing processing unit 62, the power spectrum calculation unit 64, and the frame averaging unit 66 for the waveform 92 of the target audio 32. The divider 38 divides the latter by the former, so that a PSD difference 110 is obtained.

  A mel cepstrum 120 is obtained by performing processing in the logarithm calculation unit 70, the IFFT processing unit 72, and the first mel warping unit 74 on the PSD difference 110. By executing the processing in the second mel warping unit 76, the FFT processing unit 78, and the exponent conversion unit 80 on the mel cepstrum 120, a PSD 130 in which the PSD difference 110 is smoothed is obtained. The filter parameter of the FIR 84 is set by performing processing of the IFFT processing unit 82 on the smoothed PSD 130. The frequency characteristic of the voice 42 obtained by filtering the source voice 30 using the FIR 84 set in this way is substantially equal to the frequency characteristic of the target voice 32.

  After the cepstrum is once converted to the mel scale by mel warping, the PSD difference is smoothed by removing the high frequency component by the inverse conversion. Therefore, the FIR 84 set by the filter parameters obtained in this way has characteristics close to human auditory characteristics. Further, the characteristic of the FIR 84 set in this way is not sensitive to the parameter order unlike the filter by LPC conversion. When the filter parameters are calculated so as to improve the accuracy of the filter, the sound quality is not deteriorated, and the frequency characteristics of the source sound 30 are equalized with the sound quality comparable to that of the conventional channel equivalent device 200. Can do.

[Second Embodiment]
In the apparatus of the first embodiment described above, filtering is performed by FIR. However, the present invention is not limited to that using FIR. When a filter that can be directly set by the mel cepstrum coefficient output from the first mel warping unit 74 shown in FIG. 1 is used, the configuration becomes simpler. FIG. 3 shows a block diagram of a channel equivalent device 140 according to the second embodiment of the present invention using an MLSA (mel-logarithmic spectral application) filter as such a filter.

  In FIG. 3, the same components as those in FIGS. 1 and 4 are denoted by the same reference numerals. Their functions are also the same. Therefore, detailed description thereof will not be repeated here.

  Referring to FIG. 3, channel equivalent device 140 according to the third embodiment differs from channel equivalent device 20 shown in FIG. 1 in that it includes an MLSA filter instead of cepstrum filter unit 40 shown in FIG. The MLSA filter unit 150 is included. The MLSA filter unit 150 differs from the cepstrum filter unit 40 of FIG. 1 in that a second melwarping unit 76, an FFT processing unit 78, an exponent conversion unit 80, and an IFFT processing unit 82 in FIG. This includes a MLSA filter 160 in which filter parameters are directly set by a mel cepstrum output from the mel warping unit 76.

  The operation of the channel equivalent device 140 according to the second embodiment is the same as that of the channel equivalent device 20 of the first embodiment except that the MLSA filter 160 is set by the output of the first melwarping unit 74. The same.

  The channel equivalent device 140 can also provide the same effects as the channel equivalent device 20 of the first embodiment. In addition, the MLSA filter 160 is a filter having a mel cepstrum as a parameter, and can be directly set by the output of the first mel warping unit 74. Accordingly, as compared with the channel equivalent device 20 of the first embodiment, various parts for creating FIR filter parameters from the mel cepstrum are unnecessary, and the circuit configuration is simplified.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram of the channel equivalent apparatus which concerns on 1st Embodiment. It is a schematic diagram for demonstrating operation | movement of the channel equivalent apparatus which concerns on 1st Embodiment. It is a block diagram of the channel equivalent apparatus which concerns on 2nd Embodiment. It is a block diagram of the channel equivalent apparatus which concerns on a prior art.

Explanation of symbols

20, 140, 200 channel equivalent device, 30 source speech, 32 target speech, 34 source PSD generation unit, 36 target PSD generation unit, 38 division unit, 40 cepstrum filter unit, 50, 60 speech frame detection unit, 52, 62 window Multiplication processing unit, 54, 64 power spectrum calculation unit, 56, 66 frame averaging unit, 74 first mel warping unit, 76 second mel warping unit, 84 FIR, 150 MLSA filter unit, 160 MLSA filter, 210 LPC filter Part

Claims (4)

Means for calculating a power spectral density (PSD) difference between the speech to be processed and the reference speech;
Means for converting the difference into a feature space of frequency characteristics represented by a cepstrum;
An audio frequency characteristic equalization apparatus including predetermined filtering means for filtering the audio to be processed, in which a filter parameter is set using the difference represented by the cepstrum. The filtering means includes
Means for converting the differential cepstrum into a mel cepstrum;
The audio frequency according to claim 1, further comprising: an MLSA (mel-logarithmic spectral application) filter connected so as to receive the audio to be processed as an input using the difference represented by the mel cepstrum as a filter parameter. Character equalizer. The filtering means includes
PSD difference smoothing means for performing a smoothing process on the PSD difference;
The filter parameter is set by PSD smoothed by the PSD difference smoothing means, and includes an FIR (fine impulse response) filter connected to receive the speech to be processed. Equalizer for frequency characteristics of audio. The audio frequency characteristic equalizer further includes means for converting the PSD difference into a differential cepstrum having a predetermined first order,
The PSD differential smoothing means performs a first mel for warping the difference cepstrum from the first order cepstrum to a mel cepstrum having a second order smaller than the first order. Warping means;
A second melwarping conversion means for performing an inverse transformation of the melwarping by the first melwarping means and outputting a cepstrum of the inversely transformed difference to the output of the first melwarping means;
Means for outputting a PSD difference smoothed by converting the cepstrum of the inversely converted difference output from the second mel warping conversion means into a spectrum, and giving the difference as a filter parameter to the FIR filter; The equalization apparatus of the frequency characteristic of the audio | voice of Claim 3 containing.

Images