JP2558658B2 - Basic frequency analyzer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement of a device for analyzing a fundamental frequency of a voice, and in applications such as voice analysis / synthesis, voice confirmation, and high-efficiency coded transmission of voice. The present invention relates to a fundamental frequency analysis device for sound used.

[Conventional technology]

There are many methods of extracting the fundamental frequency of the speech, such as a method using a cepstrum and a method using the autocorrelation function of the speech waveform itself, but as a conventional technique corresponding to the present invention, a residual waveform obtained as a result of linear prediction analysis. A method of using the autocorrelation function of will be described.

FIG. 2 shows how to obtain the fundamental frequency according to the prior art.
Further, FIG. 3 schematically shows the data until the normalized autocorrelation function is obtained.

First, the input voice waveform is a temporally continuous waveform whose state changes over time (changes from voiced to unvoiced, or the articulation style changes). By 21, a certain period (for example, 20 msec) is taken out. Then, by applying a window function 22 such as a Hamming window or a Hanning window to weaken the amplitude of the wave at the edge of the window, the influence at the time of the cutout boundary of the window is reduced. Then, the linear prediction analysis 23 is performed on the assumption that the state of the system that generates the speech waveform is stationary during the window length. By the linear prediction analysis, the residual waveform corresponding to the sound source information and the coefficient corresponding to the sound path transfer function are obtained. The residual waveform is almost random noise in the case of no voice, and the impulse train corresponding to the vibration of the vocal cords is obtained in the case of voice (FIG. 34 in FIG. 3).

To find the fundamental frequency, it suffices to find the cycle of repetition of this impulse sequence, so the normalized autocorrelation function of the residual waveform (autocorrelation function divided by power, delay time τ = τ =
When the function is normalized so that the function value becomes 1 when 0), the normalized autocorrelation function delayed by an integral multiple of the period of the fundamental frequency ( _defined as τ _p ) exhibits a maximum value. To obtain the fundamental frequency, the fundamental frequency may be calculated by finding this maximum value from the normalized autocorrelation function of the residual waveform and calculating the reciprocal of τ _p . Actually, there is a case of unvoiced sound or a case of no sound (a period in which neither voiced sound nor asexual sound is produced). Therefore, the processing is performed as described below.

In the case of the normalized autocorrelation function of the residual waveform of unvoiced sound, it takes a relatively small value except that the maximum value 1 is taken when the delay time τ is 0. In the case of the normalized autocorrelation function of the residual waveform of voiced sound, τ = τ _p , τ = 2τ in addition to the maximum value of 1 at τ = 0.
_The maximum value is reached with _p , .... In addition, a value of τ that is not related to the period τ _{p of the} fundamental frequency often exhibits a maximum value. In addition, when there is no sound, originally no signal should appear in the voice waveform, but since there is usually noise (especially hum of 50 or 60 Hz from the power supply), the noise causes a normalized autocorrelation function. As in the case of voiced sound, it often has local maximum values. However, since the noise is smaller than the voice waveform, the power of the residual waveform can be discriminated by its small value. Actual processing including discrimination between voiced, unvoiced, and silent is performed as follows.

First, a threshold value θ _p is provided, and when the residual waveform power obtained by the length of the window does not exceed θ _p , it is determined that there is no sound (the fundamental frequency at this time is “none”). If it is not silent, a threshold value θ _v is set next, and if the maximum value other than τ = 0 of the normalized autocorrelation function of the residual waveform does not exceed θ _v , it is determined to be silent (also in this case The frequency is "none". It is the case of voiced sound when it is neither silent nor unvoiced, and the value of τ that takes a maximum value exceeding θ _v except near τ = 0 of the normalized autocorrelation function of the residual waveform is, in order from the largest local maximum value. The first candidate, the second candidate, the third candidate, ... Of the cycle of the fundamental frequency, and the reciprocal of the cycle are taken as the first candidate, the second candidate, the third candidate of the fundamental frequency.
26).

The above candidates for the fundamental frequency are obtained for each window position while gradually shifting the window position (for example, 10 msec). FIG. 4 schematically shows how the window positions are shifted. FIG. 5 shows the states of the candidates of the fundamental frequency obtained for each window position up to the second candidate. In the figure, it is the first candidate, and is the second candidate. The fundamental frequency finally obtained is basically the frequency of the first candidate, but the frequency of the second candidate and the third candidate is the window of each window depending on the connection with the candidate of the fundamental frequency at the front and rear window positions. If the temporal change pattern of the fundamental frequency of the position continues smoothly, that is the fundamental frequency (see reference numeral 27 in FIG. 2). The fundamental frequency at the window position N is selected from the candidates as follows, for example.

First, when the first candidates of the fundamental frequencies at the window positions of N-2, N-1, N, N + 1, and N + 2 are obtained, they are 128, 104, 121, 107, and 108 Hz in order. If these are arranged in descending order, they are 128, 121, 108, 107, 104 Hz, and the third or central value is 108 Hz. Window position N
In, the value closest to this central value is selected from the candidates as the fundamental frequency. That is, since the second candidate 106 Hz is closer to 108 Hz than the first candidate 121 Hz, the second candidate is finally set as the fundamental frequency. This processing is performed on the candidates of the fundamental frequency at all window positions to determine the fundamental frequency at each window position.

[Problems to be solved by the invention]

In the prior art, an analysis was performed by setting a window and assuming that the voice generation system is in a steady state within that period. Therefore, there are the following problems.

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

(2) The window length is not always optimal. The longer the window, the longer the period assuming the steady state (actually, it is non-steady), and the fundamental frequency cannot be extracted where the fundamental frequency changes rapidly. On the contrary, if the window is short and is smaller than twice the period of the fundamental frequency, the fundamental frequency may not be extracted (which is likely to occur in the case of low voice).

The variation range of the fundamental frequency reaches 2 octaves even for the same person in conversation, so it is difficult to always be optimal with a finite and fixed window length. To improve the analysis accuracy, the window length should be adaptive. It is desirable to change to.

[Means for solving problems]

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

Reference numeral 1 denotes a portion for continuously performing linear prediction analysis (here, consideration is made to include partial autocorrelation analysis equivalent to linear prediction analysis or analysis by a modified method thereof), and outputs a temporally continuous residual waveform. A low-pass filter 2 emphasizes the low-frequency component of the residual waveform. 3 indicates the use of a semi-infinite window function with exponential decay in the past direction. 4 indicates that the autocorrelation function is continuously calculated.
Although 3 and 4 are shown separately for the sake of clarity, they can be processed at once as described in the embodiment. Reference numeral 5 is a portion for setting a threshold for the maximum value of the autocorrelation function and the power of the residual waveform and advancing the process of finding the candidate of the fundamental frequency only for the voiced sound. Reference numeral 6 is a portion for obtaining a candidate for the fundamental frequency from the delay time τ that takes the maximum value of the autocorrelation function. Reference numeral 7 is a portion for selecting a fundamental frequency from candidates for the fundamental frequency.

[Action]

FIG. 6 conceptually shows the process of processing a voice waveform according to the present invention. Reference numeral 61 is an input voice waveform. The continuous linear prediction analysis was performed by 1 and the residual waveform was obtained.
62. Further, when the high-pass component is attenuated and the low-frequency region is emphasized by the low-pass filter 2, a waveform such as 63 is obtained. Reference numeral 64 indicates the portion of the residual waveform that is multiplied by the window function. Reference numeral 65 indicates a semi-infinite length window function having an exponential decay in the past direction. Multiplying 64 and 65 at each sample time gives a waveform 66 with a large wave size at the beginning of the window but a smaller wave size as the window ages.
This means that the present value is weighted with respect to the past value. Further, the autocorrelation function of 67 is calculated by calculating the autocorrelation function for 66 waveforms.
In FIG. 3, it is shown in a laterally enlarged state after being normalized for easy understanding.

Whether or not the voice is present can be discriminated based on the power of the residual waveform and the value of the normalized autocorrelation function as described in the related art. From the autocorrelation function, the values of the delay times τ ₁ , τ _2, ... Which show the maximum value can be used as the candidates of the cycle of the fundamental frequency, and the candidates of the fundamental frequency can be obtained by using the autocorrelation function. Select from among. FIG. 6 shows the result 66 obtained by multiplying the residual waveform by a window function of a semi-infinite length that decays exponentially for the purpose of conceptual explanation. However, 66 is not required as shown in the embodiment. The autocorrelation function can be obtained directly.

With the above processing, the fundamental frequency when the window start position is at the sampling point n can be obtained. The time-varying pattern of the fundamental frequency is obtained by finding the fundamental frequency at each position while sequentially shifting the start position of the window.

〔Example〕

First, as an example of a linear prediction analysis for obtaining a continuous residual waveform, a method for obtaining a predicted residual waveform by a lattice method is shown in FIG. The lattice method is described in Saito and Nakata: "Basics of Speech Information Processing", Ohmsha (1981), pages 110-112. Here, the conventional grid method is modified so that it can be processed continuously in time.

In FIG. 7 (a), ▲ k ⁽ⁱ⁾ _t ▼ (i = 1, ..., p) is a partial correlation coefficient at a certain sample time t, and is calculated by the continuous cross-correlation calculator C _i at each sample time. Ask. The forward prediction residual obtained as the output of the i-th stage at a certain sample time t is ε ⁽ⁱ⁾ _{f, t} ▼, and the backward prediction residual is ▲ ε ^(i).
_{If b, t} ▼, the forward prediction residual of the i + 1th stage (Fig. 7 (b)) is The backward prediction residual is Calculate with. The backward prediction residual is actually at the previous sample time t ′ = t−1, By calculating and passing that value through the delay element,
It is output as ▲ ε ^{(i + 1)} _{b, t} ▼ at the sampling time t. The forward prediction residual and the backward prediction residual of the first stage (i = 0) are (Input waveform at the time of the previous sample). Also, k ^{(i + 1)} _t ▼ is Calculate by here, Means time average. Here, since the time averages obtained by weighting the past values are continuously calculated, the time averages are calculated by applying an exponentially decaying window. As an example of the window, when the current time is set to 0 and the value of the window function for the time traced back by jτ time is W (jτ), the function of W (jτ) = exp (-jτ / Tc) is used. . Where τ is the sampling period and Tc is a parameter that determines the decay rate of the window function.
Select about msec.

The time average ▲ ▼ of the sample values y _t obtained in time series at the sampling time t is Defined by Time average obtained at the last sample time t-1 If you use ▲ ▼ can be calculated by the recurrence formula. By using this recurrence formula, Can be calculated efficiently. The residual waveform input to the low-pass filter of FIG. 1 and FIG. 2 uses the forward prediction residual ▲ ε ^{(i + 1)} _{f, t} ▼ obtained as the output of the p-th stage.

Next, the low pass filter 2 will be described. Although the fundamental frequency can be obtained without a low-pass filter, the fundamental frequency component contained in the residual waveform can be emphasized by using a low-pass filter with a cutoff frequency characteristic of about 500Hz, which improves analysis accuracy. You can

An example of a semi-infinite window function with exponential decay towards the past direction of 3 is then W ₁ (nτ) = A ₁ exp (−nτ / T ₁ ) (n ≧ 0) ( However, A ₁ = 1 and τ is a sampling period.) Is a first-order exponential function. The value of T ₁ is a parameter that determines the decay rate of the window function.
ec, a value of about 10 msec is used for women.

Let σ (n, k) be the value of the autocorrelation function at a sample point n away from the sample by k samples, and use the value σ (n−1, k) of the autocorrelation function at the immediately preceding sample point n−1. The autocorrelation function at the sample point n can be calculated by the following recurrence formula.

σ (n, k) = α ² σ (n−1, k) + α ^k _xn x _nk where α = exp (−τ / T ₁ ) x _n is the sample value of the residual waveform τ is the sampling period k = 0,1,2, ... The initial value, that is, the state before x _n is input is σ
(0, k) = 0 (k = 0,1,2, ...). The window function with exponential decay in this application can thus also be implemented by a recurrence formula rather than calculating the exponential function itself each time.

By using the above recurrence formula, the autocorrelation function at each sample point can be obtained with a relatively small amount of calculation. In the case of k = 0, that is, σ (n, 0) represents the power of the residual waveform after applying the window. The normalized autocorrelation function is represented by σ (n, k) / σ (n, 0) (k = 0,1,2, ...). The range of k to be actually calculated is as follows, where k = 0, 1, 2, ..., K _max . If the lower limit of observed may fundamental frequency is F _min, the maximum point of the autocorrelation function according to F _min appears at the delay time _{k max τ = 1 / F min} . Therefore, if the autocorrelation function is calculated up to k _max = ceil (1 / (τ F _min )), the maximum point can be found. However, ceil (x) represents the smallest integer exceeding x. F _min is set to, for example, about 70 Hz for a normal utterance of a male speaker.

The processes 5 and 6 can be performed in the same manner as the conventional technique.
The selection of the fundamental frequency 7 can be performed in the same manner as the conventional technique except that the fundamental frequency candidates are obtained for each sample point and can be compared more finely in time.

As can be seen from the above description, σ (n, k) is obtained for every sample point n, that is, continuously in time, so that the fundamental frequency can also be output for every sample point n. However, in terms of using the fundamental frequency, for every sampling point (for example, sampling frequency
In many cases, it is not always necessary for 0.1 msec for 10 kHz). For example, it is sufficient to know every 10 msec. For that purpose, it is possible to use the basic frequency obtained from 7 by thinning it out every 10 msec, or to carry out the processing after 3 or 4 every 10 msce.

In addition, as another example of a window function of semi-infinite length having an exponential decay toward the past direction of 3, a multi-order exponential window is also possible, for example W ₂ (nτ) = A ₂ nτ exp (1-nτ / T ₂ ) where A ₂ = 1 / T ₂ or W ₃ (nτ) = A ₃ (nτ) ² exp (2-nτ / T ₃ ) where A ₃ = 1 It is also possible to calculate by the formula / (2T ₃ ) ² .

In the case of W ₂ (nτ) and W ₃ (nτ), it becomes a little complicated like W ₁ (nτ), but the autocorrelation function can be calculated by the recurrence formula, so that the calculation can be performed efficiently.

FIG. 8 shows the window function shapes of W ₁ (nτ), W ₂ (nτ), and W ₃ (nτ) at 81, 82, and 83, respectively. not to mention,
W ₂ (nτ) and W ₃ (nτ) that decay in the past direction are exp (1-nτ / T ₂ ), exp (2-n), respectively.
This is due to the exponential term τ / T ₃ ). In Fig. 8, the values of T ₂ and T ₃ are the time when W ₁ (nτ) becomes 0.5 and the value of W _{2
(Nτ), W 3 (nτ} ) so as to match the time of taking the maximum value, specifically, it is shown as _{T 2 = T 1 ln2 T 3} = T 2/2.

〔The invention's effect〕

In the past, all problems that could not be avoided because of the use of finite length windows or that required special consideration are avoided. That is, depending on the position of the window, the accuracy of fundamental frequency analysis is deteriorated, and it becomes unnecessary to strictly control the length of the window.

Further, in the conventional technique, the analysis was performed on the assumption of stationarity within the range of the window, but in the present invention, the analysis is weighted by the current voice waveform as compared with the past voice waveform,
Improvement in analysis accuracy can be expected because stationarity in the window range is not required.

Also, since a function that exponentially decays is used for the window,
The autocorrelation function at the next sample point can be calculated by a relatively simple process using the autocorrelation function at the immediately preceding sample point.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of how to obtain a fundamental frequency according to the present invention,
FIG. 2 is an explanatory diagram of how to obtain a fundamental frequency according to the conventional technique, FIG. 3 is an explanatory diagram of a processing process until obtaining a normalized autocorrelation function according to the conventional technique, and FIG. 4 is a diagram showing movement of a window position. Explanatory drawing, FIG. 5 is explanatory drawing of the candidate of a fundamental frequency, 6th.
FIG. 7 is a schematic explanatory diagram of the processing up to the calculation of the autocorrelation function according to the present invention. FIG. 7A is an overall diagram of the lattice partial correlation calculation circuit. FIG. The figure is a schematic of a window function of semi-infinite length with exponential decay towards the past. 1 ... Continuous linear prediction analysis processing, 2 ... Low-pass filter, 3 ... Processing by window function of semi-infinite length having attenuation of exponential function toward past direction, 4 ... Continuous calculation of autocorrelation function, 5, 25 ... Voiced voice discrimination processing by threshold value, 6, 26
…… Basic frequency candidate selection processing, 7, 27 …… Basic frequency selection processing, 21 …… Short-time waveform extraction processing by windowing, 22 …… Window function multiplication processing, 23 …… Linear prediction analysis , 24, ... Autocorrelation function calculation processing, 31, 61 ... Speech waveform, 32 ... Window length speech waveform, 33 ... Window function multiplied waveform, 34 ... Residual waveform, 35 ... … Normalized autocorrelation function,
62: Residual waveform that is continuous in time, 63: Residual waveform after passing through a low-pass filter, 64: Residual waveform of a portion that is multiplied by a window function of semi-infinite length, 65: Semi-infinite length Window function,
66 ... Windowing result, 67 ... Autocorrelation function, 81 ... Outline of window function represented by first-order exponential function, 82 ... Outline of window function represented by second-order exponential function, 83 …… A third-order exponential window function.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Mikio Yamaguchi 1-3-1 Shimaya, Konohana-ku, Osaka Sumitomo Electric Industries, Ltd. (56) Reference JP-A-52-26107 (JP, A) Kaisho 61-187000 (JP, A)

Claims (5)

(57) [Claims] 1. A method for obtaining a residual by continuously performing linear prediction analysis with a speech waveform as an input, and a residual having a semi-infinite length window function having an exponential decay in the past direction. It has a means for obtaining the autocorrelation function and a means for obtaining the period of the fundamental frequency from the maximum point of the autocorrelation function of the residual, and continuously obtains the residual from the speech waveform, and calculates the exponential function toward the past direction. A fundamental frequency analysis device, characterized in that an autocorrelation function of a residual is obtained by applying a window function of a semi-infinite length having attenuation, and a fundamental frequency is obtained from a maximum point of the autocorrelation function. 2. The fundamental frequency analyzer according to claim 1, wherein the means for continuously performing linear prediction analysis to obtain the residual error is a continuous calculation of partial correlation by a lattice method. Fundamental frequency analyzer. 3. The fundamental frequency analyzer according to claim 1 or 2, wherein the window function W (nτ) is A ₁ exp (−nτ / T ₁ ) (A ₁ and T ₁ are constants. Τ represents a sampling period, and n is a non-negative integer.) The fundamental frequency analyzer. 4. The fundamental frequency analyzer according to claim 1 or 2, wherein the window function W (nτ) is A ₂ nτexp (1-nτ / T ₂ ) (A ₂ , T ₂ are A constant, τ represents a sampling period, and n is a non-negative integer.) The fundamental frequency analyzer. 5. The fundamental frequency analyzer according to claim 1 or 2, wherein the window function W (nτ) is A ₃ (nτ) ² exp (2-nτ / T ₃ ) (A ₃ , T ₃ is a constant, τ is a sampling period, and n is a non-negative integer.).