JPS63124100A - Fundamental frequency analyzer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an improvement of a device for analyzing the fundamental frequency of speech, and is useful in applications such as speech analysis and synthesis, speech confirmation, and high-efficiency coding transmission of speech. This article relates to a voice fundamental frequency analysis device that can be used.

[Conventional technology]

There are many methods for extracting the fundamental frequency of speech, such as a cepstrum method and a method that uses the autocorrelation function of the speech waveform itself.However, as a conventional technique corresponding to the present invention, a residual waveform obtained as a result of linear predictive analysis is used. We will explain how to use the autocorrelation function of

FIG. 2 shows how to find the fundamental frequency according to the prior art, and FIG. 3 schematically shows the data up to finding the normalized autocorrelation function.

First, the input speech waveform is a temporally continuous waveform whose state changes over time (changes from voiced to unvoiced, articulatory style changes, etc.), but this is used to extract short-term waveforms. 21, by applying a window function such as a zero-order Hamming window or Hanning window that extracts only a certain period (for example, 20 m5ec), the amplitude of the wave at the edge of the window is weakened. Reduce the impact at the point in time. Then, linear predictive analysis 23 is performed on the assumption that the state of the system that generates the speech waveform is stationary during the length of the window.The residual waveform corresponding to the sound source information and the vocal tract transfer function The coefficient corresponding to is obtained. In the case of unvoiced sounds, a residual waveform similar to random noise is obtained, and in the case of voiced sounds, an impulse train corresponding to the vibration of the vocal cords is obtained (Fig. 3, 34).
.

To find the fundamental frequency, it is sufficient to find the repetition period of this impulse train, so the normalized autocorrelation function of the residual waveform (autocorrelation function divided by the power, delay time τ is τ - 0
, the normalized autocorrelation function delayed by an integer multiple of the period (τ,) of the fundamental frequency shows a maximum value, and the fundamental frequency is To find this, the fundamental frequency can be calculated by finding this maximum value from the normalized autocorrelation function of the residual waveform and calculating the reciprocal of τ. There may be cases where the period is not shown, so the process is performed as explained below.

In the case of the normalized autocorrelation function of the residual waveform of an unvoiced sound, it takes relatively small values except for the maximum value of 1 when the delay time τ is 0. In the case of the normalized autocorrelation function of the residual waveform of a voiced sound, in addition to taking the maximum value 1 at τ=0, τ−τ2, τ−2τ2
The maximum value is reached at , . . . . Further, even when the value has no relation to the period τ of the fundamental frequency, it often shows a maximum value.

Also, when there is no sound, no signal should normally appear in the audio waveform, but there is usually noise (especially 50% noise from the power supply).
or a 6 GHz hum), the normalized autocorrelation function often shows a maximum value periodically due to the noise, as in the case of voiced sounds. However, since noise is smaller than the voice waveform, it can be determined by looking at the power of the residual waveform, based on its small value. Actual processing including voiced, unvoiced, and silent discrimination is performed as follows.

First, a threshold value θ is set, and when the residual waveform power determined by the length of the window does not exceed θ, it is determined that there is no sound (
(The fundamental frequency at this time is "none"), and if there is no silence, then set a threshold θ, so that the maximum value of the normalized autocorrelation function of the residual waveform other than around τ−0 exceeds θ. If there is no sound, it is determined to be unvoiced (the fundamental frequency is set to "none" in this case as well).If it is neither silent nor unvoiced, it is a voiced sound, and the normalized autocorrelation function of the residual waveform is outside the vicinity of τ = 0. The τΦ values that take a maximum value exceeding 09 are taken as the first candidate, second candidate, third candidate, etc. for the period of the fundamental frequency in order from the one with the largest maximum value, and the reciprocal of the period is taken. Fundamental frequency first candidate, second candidate, third candidate, etc.
・(code 26).

The above fundamental frequency candidates are determined for each window position while shifting the window position little by little (for example, by 10 m5ec). Figure 4 schematically shows how the window position is shifted, and Figure 5 shows the fundamental frequency candidates found for each window position up to the second candidate. , ■
is the second candidate. The fundamental frequency finally determined is basically the frequency of the first candidate, but the second candidate,
If the frequency according to the third candidate has a smoother time-varying pattern of the fundamental frequency at each window position, then that is set as the fundamental frequency (see reference numeral 27 in Figure 2).The fundamental frequency at window position N is, for example, as follows: Select from the candidates as follows.

First, when finding the first candidate for the fundamental frequency at each window position of N-2, N-1, N, N+1, and N+2,
They are 128, 104.121.107.108Hz in order. Arranging these in descending order is 128.121.10
8.107.104Hz, and the third or middle value is 108Hz. At window position N, the value closest to this central value is selected from the candidates and set as the fundamental frequency.

In other words, the second candidate 106Hz is better than the first candidate 121Hz.
Since it is closer to 108 Hz than the second candidate, the second candidate is finally selected as the fundamental frequency. This process is performed on candidates for fundamental frequencies at all window positions to determine the fundamental frequency at each window position.

[Problem that the invention seeks to solve]

In the prior art, analysis is performed by setting a window and assuming that the speech generation system is in a steady state within that period. Therefore, the following problems arise.

+1) Since the position of the window with respect to the audio waveform is not always optimal, the accuracy of fundamental frequency analysis may deteriorate depending on the position of the window.

(2) The window length is not always optimal. The longer the window is, the more it is assumed to be in a steady state (in reality, it is unsteady, so the period becomes longer, and it becomes impossible to extract the fundamental frequency where the fundamental frequency changes rapidly. Conversely, the shorter the window, the longer the period becomes. , when it becomes smaller than twice the period of the fundamental frequency (
It may not be possible to extract the fundamental frequency (which tends to occur with low voices).

The range of change in the fundamental frequency reaches two octaves even for the same person during conversation, so it is difficult to always achieve the optimum with a finite and fixed window length.In order to improve analysis accuracy, the window length must be adjusted adaptively. It is desirable to change it to

[Means for solving problems]

FIG. 1 shows a block diagram of processing according to the present invention.

1 is a part that continuously performs linear predictive analysis (here, we will consider analysis including partial autocorrelation analysis equivalent to linear predictive analysis or analysis using modified methods thereof), and outputs a temporally continuous residual waveform. 2 is a low-pass filter that emphasizes low frequency components of the residual waveform.

3 indicates that a semi-infinite window with exponential decay is used; 4 indicates that the autocorrelation function is continuously calculated;
3 and 4 are shown separately to make the explanation easier to understand, but they can be processed at the same time as described in the embodiment. 5 is a part that sets dark values for the maximum value of the autocorrelation function and the power of the residual waveform, and proceeds with the process of finding fundamental frequency candidates only for voiced sounds. Reference numeral 6 is a part for determining fundamental frequency candidates from the delay time τ at which the autocorrelation function takes the maximum value. 7 is a part for selecting a fundamental frequency from fundamental frequency candidates.

[Effect]

FIG. 6 conceptually shows the processing progress of a speech waveform according to the present invention. Reference numeral 61 indicates an input speech waveform.

62 is obtained by performing continuous linear predictive analysis using 1 to obtain a residual waveform. Furthermore, low pass filter 2
By attenuating the high frequency components and emphasizing the low frequency range, a waveform like 63 is obtained.

64 indicates the portion of the residual waveform that is combined with the window function,
65 indicates a window function of semi-infinite length that decays exponentially. By multiplying 64 and 65 together for each sample time, a waveform 66 is obtained in which the wave size is large at the beginning of the window, but the wave size is reduced as time passes. This means that the current value is weighted relative to the past value.

Further, the autocorrelation function of 67 is obtained by calculating the autocorrelation function for the 66 waveforms. In FIG. 3, the image is normalized and enlarged in the horizontal direction for clarity.

It is possible to determine whether a voice is voiced or not based on the power of the residual waveform and the magnitude of the normalized autocorrelation function, as shown in the prior art. From the autocorrelation function, the value of the delay time τ1, τ, ..., which indicates the maximum value, can be used as a candidate for the period of the fundamental frequency, as in the conventional technology. Select from frequency candidates. 6th
The figure shows the result 66 obtained by multiplying the residual waveform by a semi-infinite window function that decays exponentially for conceptual explanation, but as shown in the example, it is not necessary to obtain 66 directly Correlation functions can be found.

Through the above processing, the fundamental frequency when the window start position is at sample point n can be found. By sequentially shifting the starting position of the window and finding the fundamental frequency at each position, the temporal change of the fundamental frequency can be found.

[Example] First, as an example of linear prediction analysis to obtain a continuous residual waveform, a method for obtaining a predicted residual waveform by calculating the grid method is shown in Fig. 7.The grid method is described by Saito, Chunichi: There is an explanation on "Fundamentals of Speech Information Processing", Ohmsha (1981), pages 1) 0-1) 2. Here, the conventional lattice method is modified so that it can be processed temporally continuously.

In FIG. 7(a), k z"(i"' 1 +
``'+ p) is the partial correlation coefficient at a certain sample time t, which is obtained by the continuous cross-correlation calculator C5 for each sample time.The forward prediction residual obtained as the output of the first stage at a certain sample time t If the difference is εI′1 and the backward prediction residual is ε, jtl, then the 4th+1st stage (Fig. 7(b)
), the forward prediction residual for l) ε8t-
Calculate by 5. The backward prediction residual is actually at the previous sample time t'-t-1, ε;), 1(i:+
16:l.

By calculating and passing that value through the delay element,
At sample time t, it is output as εvv. Shodan (
The forward prediction residual and backward prediction residual of i=o) are ε: * 'c-Xt (input waveform) ε;, "L-x
L-1 (the input waveform at the previous sample time), and 21) 1) is calculated by: Here, means time average. Here, in order to continuously calculate the time average that weights past values, the time average is calculated by applying an exponentially decreasing window. When is set to 0 and the value of the window function at the time when jτ waiting time has increased in the past is set to W(jτ), the following function is used: W(jτ)mexp(−jτ/Tc). Here, τ is the sampling period, and Tc is a parameter that determines the speed of decline of the window function, which is selected to be about 10 to 2 O-sec.

The time average y at sample time t of the sample values y obtained in time series is defined as yt-Σyt-jW (jτ). Using the time average yt-1 obtained at the immediately preceding sample time t-1, Yt -yz-+ eXp(f/TJ 77 can be calculated using the recurrence formula Yc.Using this recurrence formula Then, εJ!tεJ!-1εt+'t, εb't can be calculated efficiently.The residual waveform input to the low-pass filter in FIG. Use !d.

Next, the second low-pass filter will be explained.

The fundamental frequency can be determined without a low-pass filter, but by using a low-pass filter with a cutoff frequency characteristic of about 500 Hz, the fundamental frequency component contained in the residual waveform can be emphasized, increasing the accuracy of analysis. I can do it.

Next, an example of a semi-infinite window function of 3 is Wt(n r)
-At exp (-n f/T+) (n≧O) (where A1-1, and τ is the sampling period.

) is a first-order exponential function expressed as The value of T is a parameter that determines the speed of decline of the window function, for example, 1 for men.
For women, a value of about 10m5ec is used.

If the value of the autocorrelation function at sample point n with a place a few samples away is σ(n, k), then using the value of the autocorrelation function at the immediately preceding sample point n-1 σ(n-1, k), The autocorrelation function at sample point n can be calculated using the following recurrence formula.

σ(n,k)mα”r(n-1+k)+α”x,xll
−.

However, α= exp (-τ/'r I)xl is the sample value τ of the residual waveform, where the sampling period is =o, 1.2. ....

It is. The initial value, that is, the state before x7 is input is σ
(0, k)-0 (k-0, 1, 2, . . . ).

By using the above recurrence formula, the autocorrelation function at each sample point can be obtained with a relatively small amount of calculation. For k-Q, i.e. σ(n.

0) represents the power of the residual waveform after applying the window. The normalized autocorrelation function is e (n, k)/a (n, 0) (k=o, 1.2.-
, ). The range of k that should actually be calculated is k=0
．． 1, 2. ..., □8, it is as follows. If the lower limit of the fundamental frequency that can be observed is F sin, the maximum point of the autocorrelation function due to F + mia appears at the delay time klI□τ” 1 / F−t, l. Therefore, ksax = cet1 (1/ ( r F++
If you calculate the autocorrelation function up to ++++ )), you can find its maximum point. However, ceiHxl represents the smallest integer that exceeds X. F sin is set to, for example, about 70 Hz for normal speech by a male speaker.

The processing in 5.6 can be performed in the same manner as in the prior art. 7
The selection of the fundamental frequency can be performed in the same manner as in the prior art, except that a candidate for the fundamental frequency' is obtained for each sample point, and a more detailed comparison can be made in terms of time.

As can be seen from the above explanation, since σ(n, k) is obtained for every sample point n, that is, continuously in time, it is also possible to output the fundamental frequency for every sample point n. However, from the point of view of using the fundamental frequency, it is often not necessary to check every sample point (for example, if the sampling frequency is 10 kHz, Q, every 1 m5 ec), and it is sufficient to find out every 10 m5 ec, for example. To do this, we need to change the fundamental frequency obtained from 7 to 10
It is also possible to thin out and use it every +*sec, or to perform the processing after 3 or after 4 every 1Qmsec.

Furthermore, to give an example of a window with a semi-infinite length, a multi-order exponential window is possible, for example, Wz(n r) = A
zn T eXp (1-nτ/T2) However, Az=
1/T wealth, w, (n r) = As(n τ)! eXp (2-
n r /Tz) However, A! =1/ (2T2
)” can also be calculated.

In the case of Wz(nr) and W, (nr), it is possible to calculate the autocorrelation function using a recurrence formula in the same way as W and (n), although it is somewhat complicated, so it can be calculated efficiently.

Figure 8 shows W, (n r) , W, (n r) % v+
r, (n r) the form of each window function is 81.82.8
Shown in 3.

In Fig. 8, the values of T2 and T are specified so that when W, (nr) reaches 0.5, the time coincides with the time when W, (nr) and W3 (nr) take their maximum values. For, T, =7
1) It is shown as n2 Ts = Tz / 2.

〔Effect of the invention〕

Conventionally, this was unavoidable because a window of finite length was used.
Alternatively, all problems that required special consideration can be avoided. That is, it is no longer necessary to cause the fundamental frequency analysis accuracy to deteriorate depending on the position of the window, and to strictly control the length of the window.

Furthermore, in the conventional technology, analysis was performed assuming stationarity within the range of the window, but in the present invention, the analysis places more weight on the current audio waveform compared to the past audio waveform. Since this method does not require any specificity, it can be expected to improve the accuracy of analysis.

Furthermore, since an exponentially decreasing function is used for the window, the autocorrelation function at the next sample point can be calculated by relatively simple processing using the autocorrelation function at the immediately previous sample point.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of how to find the fundamental frequency according to the present invention,
Fig. 2 is an explanatory diagram of how to obtain the fundamental frequency using the prior art, Fig. 3 is an explanatory diagram of the processing process up to obtaining the normalized autocorrelation function using the prior art, and Fig. 4 is an explanatory diagram of how to obtain the normalized autocorrelation function using the prior art. 5 is an explanatory diagram of fundamental frequency candidates, FIG. 6 is a schematic explanatory diagram of the process up to autocorrelation function calculation according to the present invention, and FIG. 7(a) is a lattice partial correlation diagram. The overall diagram (b) of the calculation circuit is a partial circuit diagram thereof, and FIG. 8 is a schematic diagram of the window function. 1... Continuous linear predictive analysis processing, 2...
・Low-pass filter, 3...Processing using a semi-infinite window that decreases exponentially, 4...Continuous calculation of autocorrelation function, 5.25...Voicing using a threshold value Determination processing, 6.26... Fundamental frequency candidate selection processing, 7.27... Fundamental frequency selection processing, 2
1... Short-time waveform extraction processing by windowing,
22...Processing of applying a window function, 23...
・Linear prediction analysis, 24, ... Autocorrelation function calculation process, 3L 61 ... Audio waveform, 32...
...Audio waveform with window length, 33...Waveform multiplied by window function, 34...Residual waveform, 35...
...Normalized autocorrelation function, 62...Temporally continuous residual waveform, 63...Residual waveform passed through a low-pass filter, 64...Semi-infinite length The window function and multiplication are the residual waveform of the part to be combined, 65... A half-infinite length window function, 66... The window multiplication is the result, 67...
・・・Autocorrelation function, 81...--Outline of the window function expressed by the following exponential function, 82......The window function expressed by the quadratic exponential function Approximate form, 83... Approximate form of a window function expressed as a third-order exponential function. Patent applicant Hiroshi Fujisaki Sumitomo Electric Industries Co., Ltd. Agent Kama 1) Text 2 Figure 1 Figure 2 Figure 5 Fundamental frequency H! Fig. 6 61 Speech waveform force 伜 W 穎 孕 y V 内 や ■ 62 Residual waveform - Gaku Y goods 1 S P warehouse - ~ ■ 66 Windowing result fundamental frequency period candidate Fig. 7 (a) ■: Addition Device I]: Delay element for 1 sample (b) Figure 8

Claims (5)

[Claims] (1) A means for obtaining a residual by continuously performing linear predictive analysis using an audio waveform as an input, and a means for obtaining an autocorrelation function by multiplying the residual by a window function of semi-infinite length that decays exponentially; means for determining the period of the fundamental frequency from the maximum point of the autocorrelation function of the residual; the residual is continuously determined from the audio waveform, and the residual is multiplied by a window function of semi-infinite length that decays exponentially. 1. A fundamental frequency analysis device characterized in that the autocorrelation function of the autocorrelation function is obtained, and the fundamental frequency is obtained from the maximum point of the autocorrelation function. (2) In the fundamental frequency analyzer according to claim 1, the fundamental frequency analysis device is characterized in that the means for continuously performing linear predictive analysis to obtain residuals is based on continuous calculation of partial correlation using a lattice method. Analysis equipment. (3) In the fundamental frequency analysis device according to claim 1 or 2, the window function W(nτ) is A_1exp(-nτ/T_1) (A_1 and T_1 are constants. τ indicates a sampling period. .n is a non-negative integer.) A fundamental frequency analyzer characterized by being represented by: (4) In the fundamental frequency analyzer according to claim 1 or 2, the window function W(nτ) is A_2nτexp(1-nτ/T_2) (A_2 and T_2 are constants. τ is the sampling period. (n is a non-negative integer.) A fundamental frequency analyzer characterized by being represented by: (5) In the fundamental frequency analyzer according to claim 1 or 2, the window function W(nτ) is A_3(nτ)^2exp(2-nτ/T_3)(A_
3.T_3 is a constant. τ indicates the sampling period. n is a non-negative integer. ) A fundamental frequency analyzer characterized by: