JPH01238698A - Voice fundamental period extractor - Google Patents

Abstract

PURPOSE:To stably extract a voice fundamental period even if there is much change in the voice fundamental period by providing plural voice analyzing means, a selective coupling section for the outputs thereof and a voice fundamental period extracting section. CONSTITUTION:A microphone 1 is led via an A/D converter 2 to the voice fundamental period extractor by a cepstrum method. A spectral analyzing section 3-1 is shortest in analysis section length and 3-n is longest. The analysis section length of the analyzing section 5-5 past a logarithmic conversion section 4 is shortest and is longest in 5-n. The adequate cepstra of the analysis section length are selected and coupled according to the frequency part in a cepstrum selective coupling section 6. The max. value of the cepstra obtd. in the selective coupling section 6 is determined and the position of said max. value is determined as the fundamental period in the fundamental period extraction section 7. The flexible follow-up to the fundamental frequencies of the heavily changing voices is enabled and the stable extraction of the fundamental frequencies is possible according to this constitution.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a method for extracting the fundamental period of speech related to speech analysis, and is capable of stably extracting the fundamental period even when there are many changes in the fundamental period of speech. The purpose of the present invention is to provide a speech fundamental period extraction device that can analyze an input speech digital signal, and includes a plurality of speech analysis means each having a different analysis section length, and for overlapping analysis sections, one of the above-mentioned analysis means. a selection combination unit that selects the output of the analysis means that can perform optimal peak detection and combines the outputs of each analysis unit; and a fundamental period extraction unit that performs peak detection corresponding to the fundamental period of speech from the output of the selection combination unit. It is constituted by comprising a section.

[Field of Industrial Application] The present invention relates to a method for extracting the fundamental period of speech related to speech analysis, and in particular, it is a method for stabilizing the fundamental period even in the case of speech whose fundamental period changes greatly. The present invention relates to a speech fundamental period extraction device capable of extracting a fundamental period of speech.

[Prior Art] FIG. 6 is a diagram showing an example of a fundamental period extraction method using the conventional cepstrum method. In the figure, 51 is a microphone that converts audio into an electrical signal, 52 is an AD converter that digitizes the audio electrical signal, and 53 is a short circuit that converts the audio digital signal into a fixed section length every several milliseconds to several tens of milliseconds. A spectrum analysis section that analyzes the interval power spectrum, 54 a logarithmic conversion section that calculates the logarithm of the power spectrum obtained by the spectrum analysis section 53, 55 a cepstrum analysis section that calculates the cepstrum from the logarithmic spectrum, and 56 a peak on the cepstrum. It represents a pitch extracting unit that extracts as the pitch (fundamental period) of the voice.

The audio signal sampled and digitized in the AD converter 52 is designated as +Xo+. Spectrum analysis section 53
In , the short-term power spectrum (P8) is calculated using Equation I and Equation (2). Here, when the sampling frequency is f5)1z and the length of the analysis interval is τ seconds, the number of samples (L) in the analysis interval can be obtained by formula (2). Equation I can be efficiently calculated using fast Fourier transform (FFT) algorithms. In formula I, j is an imaginary unit (j2--1),
exp is a natural exponential function. In formula (2), F8' represents the conjugate complex number of F.

F,,-'E ω,, Xn exp(-2π
, ink/L) (Formula I) P, =F ・F,”
(Formula ■) L −τ・f5
(2) In the logarithmic conversion unit 54, P
Calculate the logarithm of , , using 2■.

Pk'=log(P,) (Formula ■) An example of a logarithmically transformed audio power spectrum (P,'') is shown in FIG. 7 (,).

The cepstrum analysis unit 55 performs cepstrum analysis by subjecting the short-term power spectrum logarithmically transformed by the logarithmic transformation unit 54 to Fourier transformation using 2).

Cn 'X P M' exp(2πjnk/
L) (2■) Figure 7(b) shows an example of the cepstrum.

The fundamental period extraction unit 56 (pitch extraction unit) extracts the cepstrum in the range of about 2 milliseconds to about 20 milliseconds, which is the fundamental cycle of human speech (when the sampling frequency is 10 kHz, n of +Cn+ is in the range of 20 to 200). Find the maximum value of and set the position of the maximum value as the fundamental period.

In addition to the cepstrum method mentioned above, there are other fundamental period extraction methods for speech, such as the autocorrelation method and the modified autocorrelation method. It is explained in detail in ``Digital Signal Processing'' edited by the Institute of Communication Engineers and ``Mr.Speech Information Engineering'' by NTT Technology Transfer Co., Ltd.

[Problem to be solved by the invention] In speech analysis, in order to stably extract the fundamental period, it is necessary to -i include the speech waveform in the analysis interval of the fundamental period (in this method, the interval in which the short-term spectrum is calculated). It is necessary to include at least several cycles.

When analyzing stationary speech, it is possible to stably obtain the fundamental period over a wide analysis interval. However, when there are many changes in the fundamental period, if a wide analysis interval is used, the analysis result will not follow the actual changes in the fundamental period of speech. Furthermore, since the fundamental period of the actual voice changes within the analysis interval, the location of the maximum value on the cepstrano\ becomes unclear, and fundamental period extraction may become unstable. Also, when the fundamental period is long (20 milliseconds or more), such as at the end of words or sentences,
With a fixed analysis length of 30 milliseconds, only about 15 audio waveforms can be included, and the problem is that the maximum value on the cepstrum is unclear, and fundamental period extraction tends to become unstable.

Figure 8 is a diagram explaining extraction of such a fundamental period, where (a) shows a case where the fundamental period can be extracted sufficiently, and (b) shows a case where extraction of the fundamental period becomes unstable. ing.

In view of the conventional problems as described above, an object of the present invention is to provide a speech fundamental period extraction device that can stably extract the fundamental period even when the actual fundamental period of speech changes. It is said that

[Means for Solving the Problems] According to the present invention, the above objects are achieved by the means described in the claims. That is, the present invention includes a plurality of audio analysis means for analyzing an input audio digital signal, each having a different analysis section length, and for overlapping analysis sections, performing optimum peak detection among the above-mentioned analysis sections. A speech fundamental period comprising: a selection combination section that selects the outputs of the analysis means that can be used and combines the outputs of the respective analysis means; and a fundamental period extraction section that detects a peak corresponding to the fundamental period of the speech from the output of the selection combination section. It is an extraction device.

[Function] In the present invention, P4 with a short fundamental period is selected from among the analysis results of a plurality of speech analysis means each having a different analysis section length.
Key is that analysis with a short analysis interval length is selected, so high time followability can be obtained, and in the case of a long fundamental period, analysis with a long analysis interval length is selected, so stable extraction can be performed. In this method, the analysis interval length is equivalently (automatically) changed without consciously changing the analysis interval length.

To further explain the operation of the present invention, for example, when the cepstrum method is adopted as a speech analysis means, a cepstrum obtained by analysis with a short analysis interval length is selected for the low quefrency part of the cepstrum, and a cepstrum obtained by analysis with a short analysis interval length is selected for the low quefrency part of the cepstrum, and for the middle quefrency part. Select a cepstrum obtained by an analysis with a medium analysis interval length, and for a part with high quefrency, select a cepstrum obtained by an analysis with a long analysis interval length, and then combine each selected cepstrum to create a new one. A selectively combined cepstrum is created, and finally fundamental period extraction is performed on this selectively combined cepstrum.

As a result, control for automatically optimizing the analysis interval length is realized as described above.

[Embodiment] FIG. 1 is a block diagram showing the configuration of a first embodiment of the present invention, and shows an example of an apparatus for extracting fundamental periods of speech using the cepstral method. In the figure, 1 represents a microphone, and 2 represents an AD converter. Also, 3-1 to 3-n
are respectively spectrum analysis units, and each spectrum analysis unit has the same function as a conventional spectrum analysis unit, but has different analysis interval lengths, and the spectrum analysis unit 3-1 has the shortest analysis interval length, and the other The spectrum analysis sections have longer analysis section lengths in numerical order, and spectrum analysis section 3-n has the longest analysis section length. Figure 2 shows the relative relationship between each analysis section. As shown in the figure, each analysis section is such that the length of the analysis section is gradually increased on both sides of the time position where the fundamental period of the voice is not analyzed. 4-1 to 4-n
is a logarithmic conversion section, which has exactly the same function as the logarithmic conversion section of the conventional example. 5-1 to 5-n are a group of gepstrum analysis units, and each spectrum analysis unit has the same function as a conventional cepstrum analysis unit, but like the spectrum analysis unit group, each has a different analysis interval length. , the cepstrum analysis unit 5-1 has the shortest analysis interval length, and the other cepstrum analysis units have longer analysis interval lengths in numerical order,
The cepstrum analysis section 5-n has the longest analysis interval length. 6 is a cepstrano\selection connection section, and a cepstrum analysis section 5-1 to 5-5-" creates a desired selection connection cepstrum from the gepstrum group obtained from 5-1. An example of the selection connection process is shown in Fig. 3. show.

Reference numeral 7 denotes a fundamental period extraction section, which has the same function as the conventional fundamental period extraction section, but calculates the maximum value of the selectively combined cepstra l\ obtained by the cepstrum selection and combination section 6, and determines the position of the maximum value. This is the basic period.

FIG. 4 is a block diagram showing the configuration of a second embodiment of the present invention, and shows an example of an apparatus for extracting fundamental periods of speech using autocorrelation analysis.

In the figure, l and 2 are the same as in Figure 1, and 8-1
8-n is a group of autocorrelation analysis units, and each autocorrelation analysis unit performs autocorrelation analysis with different analysis interval lengths as shown in FIG. Specifically, the AD-converted audio signal is set to (×1), and the window function for removing the influence of cutting out the analysis section is (ωl1l-0+L~+ (1
- is the analysis interval length), the autocorrelation function (σ4.L) when the analysis interval length is L is calculated by the following formula (2). n is the analysis center position.

Reference numeral 9 denotes an autocorrelation selection combination unit, which has the same function as the cepstrum selection combination unit in the case of the Gebstral analysis described above.

Further, the fundamental period extraction section 10 also has the same function as in the first embodiment.

FIG. 5 is a block diagram of a third embodiment of the present invention,
An example of a device for extracting fundamental periods of speech using modified autocorrelation analysis is shown.

In the figure, 1.2 is the same as in FIG. 1 or 4, and 11-1 to 11-n are autocorrelation analysis units, which are also similar to those in the second embodiment. It is similar to the case.

12-1 to 12-n are linear prediction analysis units, and the formula ■ is calculated from the autocorrelation function (σ1, L) obtained by each autocorrelation analysis unit.
Calculate (α1.L) that satisfies the following.

The Levinson method can be used for this calculation.

(Formula ■) Next, by Formula ■, A which is the correlation coefficient of (α1.L)
Calculate parameters.

Each of 13-1 to 13-0 is a modified autocorrelation analysis section, which calculates the modified autocorrelation analysis section +W, , L) according to equation (2).

-1-Σ ^J+Lσ, -1 (Formula IX) 14 is an autocorrelation selection coupling part, which corresponds to the cepslum selection coupling part 6 in the case of cepstom analysis.

[Effects of the Invention] As described above, according to the present invention, it is possible to realize a fundamental frequency extracting device for speech that can flexibly follow the fundamental frequency of speech that changes drastically and can extract the fundamental frequency stably.

Although the method of the present invention requires a slight increase in the amount of processing compared to the conventional method, it is possible to realize an extremely excellent fundamental period extraction device without complicated algorithms.

This level of increase in processing amount is not a problem at all with recent LSI technology.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of the first embodiment of the present invention, FIG. 2 is a diagram showing the relative relationship between the analysis intervals of each spectrum analysis section, and FIG. 3 is an example of cepstrum selective combination processing. 4 is a block diagram showing the configuration of the second embodiment of the present invention, FIG. 5 is a block diagram of the third embodiment of the present invention, and FIG. 6 is fundamental period extraction using the conventional cepstral method. FIG. 7 is a diagram showing an example of the method, FIG. 7 is a diagram showing an example of a logarithmically transformed audio power spectrum and cepstrum, and FIG. 8 is a diagram explaining extraction of the fundamental period. l ・Microphone, 2...-AD converter, 3-1
~3-n-... Spectral analysis unit, 4-1 to 4-Rho... Logarithmic conversion unit, 5-1 to 5-11 Cepstrum analysis unit, 6... Cepstrum selection combination unit, 7.1
O515... Fundamental period extraction unit, 8-1 to 8-n
, 11-1 to 11-n... Autocorrelation analysis section, 9
...Autocorrelation selection combination unit, 12-1 to 12-n ・
・Linear prediction analysis section, 13-1 to 13-n ・Modified autocorrelation analysis section, 14...Modified autocorrelation selection combination section\, Agent Patent attorney Sada Igeta -. 4\X,. ＼□□ (a) Extraction of new synchronization in divided section 1 Part a (b) Diagram with two explanations between analysis y

Claims (1)

[Scope of Claims] 1. A plurality of audio analysis means each having a different analysis section length for analyzing an input audio digital signal, and for overlapping analysis sections, optimal peak detection among the above-mentioned analysis means is performed. A selection combination unit that selects the outputs of the analysis means that can perform the analysis and combines the outputs of the analysis means, and a fundamental period extraction unit that detects a peak corresponding to the fundamental period of the voice from the output of the selection combination unit. Characteristic voice fundamental period extraction device. 2. The speech fundamental period extraction device according to claim 1, wherein each of the plurality of speech analysis means comprises a respective spectrum analysis section, a logarithmic conversion section, and a cepstrum analysis section, and the selective combination section is a cepstrum selection combination section. 3. The speech fundamental period extraction device according to claim 1, wherein each of the plurality of speech analysis means is an analysis means based on autocorrelation analysis, and the selective combination section is an autocorrelation selection combination section. 4. The speech fundamental period extraction device according to claim 1, wherein each of the plurality of speech analysis means includes an autocorrelation analysis section, a linear prediction analysis section, and a modified autocorrelation analysis section, and the selective combination section is a modified autocorrelation selection combination section. 5. The speech fundamental period extraction device according to claim 1, wherein the plurality of analysis sections have the same center position and different analysis section lengths.