JPH1114672A - Method for estimating spectrum of periodic waveform and program recording medium therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly accurately estimate a spectrum, without synchronizing the spectrum with a pitch. SOLUTION: An approximately 3-5 ms minute section &tau; of an input sound waveform is sliced (S3) and subjected to DFT(discrete Fourier transform) with a Hamming window (S5). A spectrum of the &tau;, i.e., M=log(X+1) (X: a result of the DFT) is found (S6). A cumulative addition A=A+M<e> (an initial value of the A is 0) is performed (S7). The &tau; is shifted by 2-4 ms, and the slicing of the &tau; is carried out again (S8). The process is conducted for a conventional analysis section T, and S=(A/N)<1/e> (where N-1 is a count of shifts) is calculated to obtain a spectrum S (S10). This procedure is executed for each spectral component.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for estimating a spectrum of a waveform having a periodic structure such as a speech waveform at predetermined time intervals (analysis table), and a program recording medium therefor.

[0002]

2. Description of the Related Art Conventionally, in speech information processing, first, a spectrum time series is obtained from a speech waveform. In obtaining the spectrum, in the conventional short-time spectrum analysis, the voice spectrum was obtained by using the entire voice waveform included in the analysis window. Normal analysis time window width is 20 ms to 4
It is about 0 ms. Normally, the window width is twice or more than the pitch period (glottal opening / closing period, time interval from opening to opening) in phonemes (vowels, vowels,
Consonant) shorter than the length. Using such an analysis method, a harmonic structure appears in the spectrum. [For example,
Sadahiro Furui, Digital Audio Processing, Tokai University Press, 19
85]. Therefore, the spectrum shape is easily affected by the pitch period. If a speech signal of one pitch section is cut out and analyzed in synchronization with the pitch and averaged within a given section, the accuracy of spectrum estimation can be improved, but it is difficult to cut out a pitch section accurately.

[0003]

SUMMARY OF THE INVENTION It is an object of the present invention to provide a method for estimating a spectrum of a periodic signal which can estimate a correct spectrum which is not synchronized with the pitch but is not affected by the pitch period, and a program recording medium therefor.

[0004]

According to the present invention, a spectrum is analyzed in an analysis window shorter than the period T of a periodic signal, the analysis result is raised to the power of a real number, and a plurality of short-time spectra are integrated to obtain a spectrum of the period T. Is estimated. Describing the audio waveform, a short-time audio waveform is cut out at a short cycle (a minute section window width). The spectrum obtained from the short-time speech waveform is called a minute section spectrum.
Here, the short time is assumed to be a time of about one pitch or less of the voice waveform. Since such a short-time waveform does not have a temporal periodic structure, the spectrum does not show a harmonic structure. FIG. 1A shows a state of audio clipping. τ is a minute section window width (time) for obtaining a minute section spectrum, δ is a minute section shift width (time), and T is an analysis window width (time) in the conventional analysis. In order to make the effective window width the same as the conventional window width, the following number of minute section spectra may be integrated.

N = ((T−τ) / δ) +1 (1) If the following L _p norm (p-mean distance) is used as a function for integrating a plurality of minute interval spectra, various integration methods can be unified. Can be expressed as S (ω, t) = {(1 / N)} _{i = 0} ^N−1 M (ω, t + δi) ^e ｝ ^{1 / e} (2) where M (ω, t) is a minute section spectrum at time t. ,
ω represents a frequency. e (e ≠ 0) is an exponent, and when e = 1, simply represents the arithmetic mean of the minute interval spectrum. e is-
When ∞, the calculation is to find the minimum value, and when e is ∞, the calculation is to find the maximum value. The one obtained by the equation (2) is called an integrated minute section spectrum, and the spectrum analysis method using the equation (2) is called a minute section spectrum method.

[0006] Each minute section spectrum M (ω, t) is obtained by, for example, FFT. The FFT order is a power of 2, more than twice the number of frequency analysis channels, and the FFT on the waveform
The window length is set to a number exceeding the window length τ of the minute section spectrum for the first time. The waveform data of length τ multiplied by the Hamming window is inserted left-justified, and the FFT is performed after setting it to 0.
Assuming that the channel k and the time are i and the linear FFT spectrum of the minute section is P (k, i), the minute section spectrum M (k, i) used for integration is obtained by the following equation.

M (k, i) = log (1 + P (k, i)) (3) Although this is close to a logarithmic spectrum, the value is always a positive value.
Each term must be positive when calculating the Lp norm. Frequency corresponding the order of the FFT to the channel k When K is given when the sampling frequency of the audio and f _s by Equation (4).

[0008] _{ω (k) = πf s k} / (2K) (4)

[0009]

FIG. 3 shows an embodiment of the method according to the invention. First, the time pointer i in a voice section, for example, T = 30 ms is set to 0 (S1), and the content A of the buffer for storing the spectrum is cleared and initialized to A = 0 (S2). Next, a sound in a minute section, that is, a section of about τ = 5 ms (t to t + τ) is cut out (S3), and the cut out sound signal is windowed, for example, a Hamming window is applied (S4). A power spectrum X is obtained by performing a k-th order DFT (Discrete Fourier Transform) on the voice signal in the windowed section τ (S5).

With respect to the result X of the DFT, a minute section spectrum M is obtained by equation (3), that is, M = log (X + 1) (S6). This log is the natural log. The obtained small spectrum M is raised to the power of e and cumulatively added to the spectrum A stored in the storage buffer (S
7). That is, the following equation is calculated. A ← A + ^Me Next, the time pointer i is incremented by +1 and, for example, shifted by a minute section shift width δ = about 2 ms, that is, the time is set to t + δ (S8). It is determined whether the end of the voice section of T = 30 ms has been reached (S9). If it has not reached, the process returns to step S3, and the minute section shifted by δ (t + δ to t)
+ Δ + τ), and the same processing is performed thereafter.

In this way, while shifting by δ, the spectrum M of the minute section voice in section τ is obtained, and this is cumulatively added to the content A of the storage buffer, and as shown in equation (1), N When accumulating the spectrum in small sections, that is, when it reaches the end .delta.i = T = 30 ms speech segment, which is determined in the step S9, the spectral ^{_{a = Σ i = 0 N-}} 1 M e obtained by accumulating a stored content of the storage buffer Is divided by the accumulated number N, and the division result is raised to the power of 1 / e, that is, the equation (2) is calculated to obtain the minute section spectrum S (S10). In the processing shown in FIG. 2, the repetition of each frequency is omitted, and thus A, X, M, and S in FIG. 2 are vectors having elements of the number of frequency channels.

The results of a simulation experiment performed on a certain frequency channel to see how the temporal fluctuations of the values of the minute section spectrum obtained in this way are integrated are shown. That is, it is assumed that the fluctuation of the value of a certain frequency in the spectrum is a sine wave such as s (i) = 0.5−0.5 cos (4πi / N) + ε (5) 0 < i <N. ε is a minute constant for avoiding the divergence of the value.

The equation for integrating the minute interval spectrum with the Lp norm is given by: v (e) = ｛Σi _{= 0} ^N−1 s (i) ^e ｝ ^{1 / e} (6) Here, the purpose is to integrate the minute section spectra, and it is basically not a problem to find the minimum value. When low-level noise is superimposed, or when the driving sound source is unstable, it is appropriate to attach importance to a portion having a large energy. In such a case, 1 <e may be set to integrate a plurality of minute section spectra. Also, focusing on low energy parts,
For the purpose of removing sudden noise, 0 <e <1 may be set.

It is assumed that the fluctuation of the spectrum of the equation (5) is buried in a temporally constant noise ν and r (i) = max [s (i), ν] (7). That is, as shown in FIG. 1B, the horizontal axis represents time, and the vertical axis represents level, the signal s (i) changes as shown by a curve 11, and noise ν _{1 at} various levels parallel to the horizontal axis.
When ν ₂ , ν ₃ ... are superimposed, r (i) is the signal s
The larger of (i) and noise ν is r (i). FIG. 1C shows the result of calculating Equation (6) for various types of e for the signal r (i) in which the lower level portion is buried by the noise ν. It can be seen from FIG. 2 that if e is about 4 or more, the integrated spectrum level is almost constant and is hardly affected by the noise even if the noise is buried with about 1/2 of the maximum value of the spectrum. This indicates that the small interval spectrum method of the present invention is robust against noise.

In the above description, the time series of the minute section spectrum is integrated for each frequency. However, the weighted addition of the spectrum depending on the power of each minute section may be performed. That is, assuming that the power in the minute section is u (t) and the power-normalized spectrum in the minute section is Q (ω, t), the following relationship is established. u (t) = 1 / (2π) ∫M (ω, t) dω (8) Q (ω, t) = M (ω, t) / u (t) (9) In the case where weighted addition of spectra depending on is performed, the integration formula of the minute section spectrum is as follows.

S (ω, t) = { _{i = 0} ^N−1 u (t + δi) ^e Q (ω, t + δi)} / Σ _{i = 0} ^N−1 u (t + δi) ^e (10) In this case, FIG. In step 2, the small section spectrum M obtained in step S6 is converted into u (t) and Q by the relationship shown in equation (9).
(Omega, t) is divided into, in step S7, based on the equation (10) may be calculating the A = A + u ^e Q. That is, A = A + for each spectrum
u ^e may be calculated. Q (ω, t + δi) may be obtained not only by FFT but also by LPC analysis.

FIG. 4A shows an example of speech recognition using a minute section spectrum. HMM (Hidden Markov Model) [Seiichi Nakagawa: Speech Recognition by Stochastic Model,
This is the case where IEICE, 1988] is used. An input voice from the microphone 21 is passed through a low-pass filter 22 having a pass band of の of the sampling frequency (for example, 12 kHz), and is then digitized from an analog signal by an A / D converter 23 at the sampling frequency. This digital audio signal is converted into a spectrum time series by using the minute section spectrum in the minute section spectrum estimating section 24 according to the present invention. The spectrum time series of the learning speech is input to the HMM learning unit 25, where the HMM is created and stored in the HMM storage unit 26. The spectral time series of the speech to be recognized is input to the HMM recognizing unit 27, and the recognition process is performed with reference to the HMM in the HMM storing unit 26 and the list of recognized vocabulary in the vocabulary information storing unit 28, and the result is displayed on the display unit. 29 is displayed. The learning and recognition of the HMM use standard methods described in the above-mentioned references.

The processing of generating a time series of spectra using the minute section spectrum in the minute section spectrum estimating section 24 is performed as shown in FIG. 4B. First, the time pointer is set to 0 (S1), and an audio signal of T = 30 ms is cut out from the continuous audio waveform starting from the time pointer (S1).
2). A spectrum is extracted from the extracted 30 ms audio signal by the minute section spectrum method (S3). Next, the time pointer is moved for 10 ms (S4). This is a value called a frame rate or a frame period sent to the voice recognition unit 27. Next, it is determined whether or not the time pointer has reached the end of the audio signal (S5). If it has not reached the end, the process returns to step S2, and if it has reached the end, the process ends. The extraction of the minute section spectrum in step S3 is performed by the processing shown in FIG.

The present invention can be applied not only to speech waveforms but also to spectrum estimation of periodic waveforms.

[0020]

As described above, according to the present invention, in order to estimate a spectrum such as a speech waveform by integrating spectra obtained from minute sections, the value of e is selected, that is, e> 1. By doing so, it is possible to selectively integrate minute section spectra having high energy. A spectrum having high energy is obtained in synchronization with the pitch period, and a highly accurate spectrum estimation close to pitch-synchronous spectrum analysis can be performed. For this reason, the spectrum is hardly affected by the pitch period and the cutout position of the voice analysis section, and the recognition performance can be improved by applying the present invention to voice recognition.

Further, by selecting e so that 0 <e <1, it is possible to remove sudden noises with emphasis on low energy portions. As a result of a comparison between a case where a spectrum sequence is obtained by applying the present invention to speech recognition and a case where a spectrum is obtained by a conventional FFT using a window of about twice or more the pitch period, recognition of phonemes having different utterance styles was performed. As a result, the phoneme recognition rate, which was 64% in the past, could be improved to 71%. In any case where e is 0.5 to 2, the effect of the minute section spectral method can be obtained.

The window length T of one frame is 15 ms, 20 m
In any of s, 30 ms, and 40 ms, the use of the minute interval spectrum method can obtain a higher recognition rate than the conventional FFT spectrum using a long data window. This improvement is significant when recognizing sounds with different utterance styles. The improvement effect is greater for vowels. The minute section frame shift δ is preferably from 2 ms to 4 ms, and the minute section frame window length τ is preferably about 3 ms to 5 ms, that is, ５〜 of the window length T.
The length is preferably about 1/14, particularly about 5 ms for voices of the same utterance style, and about 3 ms for voices of different utterance styles.

[Brief description of the drawings]

FIG. 1A is a diagram showing an example of a relationship between a conventional spectrum analysis window T, a minute section window width τ, and a minute section shift width δ in the present invention, and B shows an output waveform of a frequency channel buried in noise. FIG.

FIG. 2 is a diagram showing a relationship between a change in noise level as a parameter and an integrated minute section spectrum.

FIG. 3 is a flowchart showing an example of a spectrum estimation method according to the present invention.

FIG. 4A is a diagram showing a functional configuration of a speech recognition apparatus to which the minute section spectrum estimating method of the present invention is applied, and FIG. 4B is a flowchart showing a procedure for generating a minute section spectrum time series.

Claims (5)

[Claims] 1. A method for estimating a spectrum for each fixed time section T of a periodic waveform, comprising: extracting a periodic waveform of the fixed time section T by a minute section τ shorter than the fixed time section. This is performed by sequentially shifting the small section shift width δ shorter than the section τ, to obtain the spectrum M of the waveform of each of the cut-out minute sections.
/ E raised to the spectrum of the above-mentioned fixed time section T, wherein a spectrum of a periodic waveform is estimated. 2. The spectrum M of the minute section is obtained by performing a discrete Fourier transform on the waveform of the minute section, and obtaining a power spectrum X obtained by the conversion result as M = log.
2. The method for estimating the spectrum of a periodic waveform according to claim 1, wherein (X + 1) (log is a natural logarithm) is calculated. 3. The minute section spectrum M is converted into a minute section power u and a spectrum Q normalized by the power u.
3. A method for estimating a periodic waveform spectrum according to claim 1 or 2, wherein the real power e is obtained for a power u in a small section by obtaining a weighted average using the e raised to the power of e. . 4. When estimating a spectrum of an audio waveform for each fixed time interval, the audio waveform of the fixed time interval T is replaced with the fixed time interval T
Cutting out shorter shorter sections τ is called smaller section τ
The shorter minute section shift width δ is sequentially shifted, and the spectrum M of the speech waveform of each of the cut-out minute sections is performed.
A recording medium on which is recorded a program for performing, by a computer, each of these minute section spectrums M raised to the power of the real number e, averaged, and further raised to the power of 1 / e to obtain the spectrum of the fixed time section T. 5. A spectrum of a voice waveform is converted into a predetermined time interval T.
When estimating each time period, the speech waveform in the fixed time interval T is
Cutting out shorter shorter sections τ is called smaller section τ
The shorter minute section shift width δ is sequentially shifted, and the spectrum M of the speech waveform of each of the cut-out minute sections is performed.
And a power u of the small sections, calculated as the product u · Q of the normalized spectral Q in power of small section, for each of these small sections spectrum M, real e raised to the power u ^e for the small sections power u A recording medium in which a program for obtaining a Q and averaging the Q to obtain the spectrum in the above-mentioned fixed time section T by a computer is recorded.