JP2956938B2 - Voice analyzer - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech analyzer, and more particularly, to a speech analyzer using a digitally converted speech signal processing technique.

[Problems to be Solved by Conventional Techniques and Inventions] Speech analysis is performed by modeling a speech generation process. Here, the voice is separated into a sound source expressing the vibration of the vocal cords and a spectrum (vocal tract characteristics) expressing the shape of the vocal tract. Generally, in this analysis,
Since the purpose is to determine the vocal tract characteristics, the sound source is considered to be other than the vocal tract characteristics. That is, a residual signal obtained by passing a voice through an inverse filter having vocal tract characteristics is used as a sound source.

In order to perform high quality speech coding and high quality speech synthesis, it is necessary to appropriately model or operate the residual signal, that is, the sound source. However, since the residual signal is spectrally whitened and is close to a random signal in waveform, it is extremely difficult to handle the residual signal.

2. Description of the Related Art Conventionally, speech analysis using digital signal processing is performed as follows in homomorphic processing, for example. That is, first, the digitally converted audio signal x (t) is subjected to a short-time Fourier transform to obtain the spectrum X (t).
(MS, ω). Then, by taking its logarithm, l
After obtaining og | X (mS, ω) |, a cepstrum is obtained by performing an inverse discrete Fourier transform on it. Thereafter, the cepstrum is processed by, for example, a 30th-order cepstrum lifter, and discrete Fourier transform and exponential transform are sequentially performed.

FIG. 4A is a characteristic diagram showing characteristics of a conventional 30th order cepstrum lifter. When the processing is performed by the lifter shown in FIG. 4A, a cepstrum having only the components of the quefrency region indicated by hatching is obtained from the cepstrum.

FIG. 4B is a characteristic diagram showing a frequency characteristic of a spectrum obtained by using the lifter shown in FIG. 4A. In FIG. 4B, the vertical axis shows the logarithmic amplitude spectrum, and the horizontal axis shows the frequency. In this figure, spectrum 4 (X (mS, ω)) obtained by short-time Fourier transform
And the envelope 5 included in the spectrum 4 (V (mS, ω))
And the other, that is, the amplitude spectrum 6 (G
(MS, ω)).

What should be noted here is that there are deep dip-shaped cut portions 61 in some parts of the spectrum 6. The reason is that a high-frequency quefrency component is included in the amplitude spectrum of the subtraction. Due to the presence of the deep cut portion 61, an audio signal obtained by converting the pitch frequency is adversely affected. Therefore, the analysis method as described above is not appropriate when converting the pitch frequency of speech.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and has as its object to obtain a speech analyzer capable of accurately performing phoneme analysis on a speech signal.

[Means for Solving the Problems] A voice analysis device according to the present invention receives a digitally converted voice signal, converts the voice signal to a spectrum signal by Fourier transform, and receives a spectrum signal from the Fourier transform means to logarithmically. A logarithmic transforming means for converting to a logarithmic spectrum signal by conversion, an inverse Fourier transforming means for receiving the logarithmic spectrum signal from the logarithmic transforming means and converting it to a cepstrum signal by inverse Fourier transforming, and a cepstrum signal from the inverse Fourier transforming means for receiving the pitch frequency. Separating means for separating the received cepstrum signal into the middle-range quefrency component near the quefrency corresponding to and the low- and high-frequency quefrency components other than the middle-range quefrency component, and the mid-range quefrency component of the cepstrum signal from the separating means. Sound source information to generate sound source information And a vocal tract information generating unit that receives the low and high frequency quefrency components of the cepstrum signal from the separating unit and generates vocal tract information.

[Operation] In the speech analyzing apparatus according to the present invention, the separating means separates the middle-range quefrency component near the quefrency corresponding to the pitch frequency in the cepstrum signal from the other low-frequency and high-frequency quefrency components. Therefore,
The effect of the harmonics of the pitch frequency can be appropriately taken out, and the sound source information and the vocal tract information can be accurately separated.

[Embodiment of the Invention] FIG. 2 is a block diagram showing a hardware configuration of a voice analyzer according to an embodiment of the present invention. Referring to FIG. 2, A / D converter 1 receives an audio signal and provides a digitally converted audio signal to an audio analyzer. The voice analysis device includes an arithmetic processing unit 2 for performing arithmetic processing of a digitally converted voice signal, and a memory unit 3 for storing a processed data signal and a necessary program. The analysis result is output from the arithmetic processing unit 2.

FIG. 1 is a block diagram showing an example of the arithmetic processing unit 2 of the speech analyzer shown in FIG. With reference to FIG. 1, the arithmetic processing unit 2 includes a short-time Fourier transform unit 21 which receives a digitally converted audio signal x (n) and outputs a spectrum X (mS, ω) by short-time Fourier transform. Receive spectrum X (mS, ω) and logarithmic spectrum log
A logarithmic spectrum converter 22 for converting into | X (mS, ω) |
An inverse discrete Fourier transform unit 23 which receives the logarithmic spectrum log | X (mS, ω) | and outputs a cepstrum by an inverse discrete Fourier transform; a component near the quefrency corresponding to the pitch frequency which receives the cepstrum; Comb-like cepstrum lifter unit 24, discrete Fourier transform units 25 and 27 for performing discrete Fourier transform on each of the separated cepstrum, and exponential conversion of the output of each of discrete Fourier transform units 25 and 27. Exponential conversion unit 26
And 28. The spectrum G (mS, ω) related to the pitch frequency is output from the exponential converter 26, and the other spectrum V (mS, ω) is output from the exponential converter 28.

The short-time Fourier transform in the short-time Fourier transform unit 21 is performed by the following equation (1).

Here, w (n) indicates the analysis window, S indicates the number of points of the shift width of the analysis window, and m indicates the number of the analysis window.

FIG. 3A is a characteristic diagram showing characteristics of a comb-like lifter used in the comb-like cepstrum lifter unit 24 shown in FIG. Referring to FIG.3A, for example, when a 512-point cepstrum is obtained by the inverse discrete Fourier transform unit 23, the comb-shaped lifter uses the 512-point cepstrum and 20 points before and after the quefrency corresponding to the pitch frequency. Is to be removed.

FIG. 3B is a characteristic diagram showing a frequency characteristic of a spectrum obtained by using the comb-shaped lifter shown in FIG. 3A. Referring to FIG. 3B, FIG. 3B shows spectrum 4 (X
(MS, ω)), a spectrum 6 related to the pitch frequency (G (ms, ω)), and the other spectrum 5 (V
(MS, ω)). The vertical axis indicates the logarithmic amplitude spectrum, and the horizontal axis indicates the frequency. When processing is performed using a comb-shaped lifter, the amplitude spectrum 6 of the subtraction becomes substantially sinusoidal, as can be seen from FIG. 3B. That is, there is no deep cut portion 61 as shown in FIG. 4B, and only the influence of harmonics of the pitch frequency can be properly extracted from the spectrum 6. Therefore, when converting the pitch frequency, the sinusoidal amplitude spectrum 6 may be modified.

In the above embodiment, the logarithmic spectrum converter 22 performs logarithmic conversion on the absolute value of the spectrum | X (mS, ω) |. However, the spectrum X (mS, ω) may be directly subjected to the complex logarithmic conversion. Good. In that case, complex exponential conversion will be performed in later exponential conversion units 26 and 28.

This technique can be widely applied in the fields of speech signal encoding, speech synthesis, and speech recognition. In other words, by extracting only the features to be converted and converting only those features, and by using as much of the features that are not related to the conversion as possible, the original ones can be used to improve the quality of the encoded speech or rule-synthesized sound. Can be achieved. Since only the parameters related to the pitch frequency of the voice are separated from the voice, the resulting residual waveform (the waveform of the portion corresponding to the pitch frequency separated from the voice) becomes impulse-like, so Not only is it suitable, but also suitable for pitch frequency conversion.

[Effects of the Invention] As described above, the present invention includes the separating means for separating the mid-frequency quefrency component corresponding to the pitch frequency and the other low-frequency and high-frequency quefrency components of the cepstrum signal, Sound source information and vocal tract information can be accurately separated.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of an arithmetic processing unit of the speech analyzer according to one embodiment of the present invention shown in FIG. FIG. 2 is a block diagram showing a hardware configuration of a voice analyzing apparatus according to an embodiment of the present invention. FIG. 3A is a characteristic diagram showing characteristics of a comb-shaped lifter used in the comb-shaped cepstrum lifter unit shown in FIG.
FIG. 3B is a characteristic diagram showing a frequency characteristic of a spectrum obtained by using the comb-shaped lifter shown in FIG. 3A. FIG. 4A is a characteristic diagram showing characteristics of a conventional 30th order cepstrum lifter. FIG. 4B is a characteristic diagram showing a frequency characteristic of a spectrum obtained by using the lifter shown in FIG. 4A. In the figure, 1 is an A / D converter, 2 is an arithmetic processing section, 3 is a memory section, 21 is a short-time Fourier transform section, 22 is a logarithmic spectrum transform section, 23 is an inverse discrete Fourier transform section, and 24 is a comb tooth. Cepstrum lifters, 25 and 27 are discrete Fourier transforms, and 26 and 28 are exponential transforms.

 ────────────────────────────────────────────────── ─── Continuing from the front page (72) Inventor Shinichi Tamura 5th Sanraya, Inaya, Seika-cho, Soraku-gun, Kyoto Pref. 5 Seira-cho, Seika-cho, Soraku-gun, 5th, Sanraya, Aitani, Japan A / T Automatic Translation and Telephone Laboratories Co., Ltd. (56) References JP-A-62-278598 (JP, A)

Claims (1)

(57) [Claims] 1. Fourier transform means for receiving a digitally converted audio signal and converting it to a spectrum signal by Fourier transform, and logarithmic transform means for receiving a spectral signal from the Fourier transform means and converting it to a logarithmic spectrum signal by logarithmic conversion. An inverse Fourier transform unit that receives a log spectrum signal from the logarithmic transform unit and converts the log spectrum signal into a cepstrum signal by an inverse Fourier transform, and receives a cepstrum signal from the inverse Fourier transform unit,
Separation means for separating the received cepstrum signal into a mid-range quefrency component corresponding to the pitch frequency and low- and high-frequency quefrency components other than the mid-range quefrency component, and the cepstrum signal from the separation means. A sound source information generating unit that receives a mid-range quefrency component and generates sound source information, and includes a vocal tract information generating unit that receives a low band and a high band quefrency component of a cepstrum signal from the separating unit and generates vocal tract information. Voice analysis device.