JPH0731504B2 - Pitch extractor - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pitch extraction device, and more particularly to a pitch extraction device that extracts a pitch period or pitch frequency of a voice via a parameter that expresses the repeatability of a voice waveform.

[Conventional technology]

Conventionally, various pitch extraction methods have been proposed in which the pitch period or pitch frequency (hereinafter simply referred to as the pitch period) of a voice is extracted through an intermediary variable expressing the repeatability of a voice waveform, for example, an autocorrelation coefficient. It can be said that a method that can be said to be perfect from the viewpoint of pitch search error has not been established yet. Among them, the method that can minimize the pitch search error is Dynamic Programming (Dynamin Programmin).
g, hereinafter abbreviated as DP) to obtain a desired pitch period sequence, which is described in detail in, for example, Japanese Patent Laid-Open No. 54-905, "Pitch Extractor".

[Problems to be solved by the invention]

However, the conventional extraction device using this type of DP has the following two drawbacks.

That is, the conventional DP selects the optimum path in a brute force manner for all the autocorrelation coefficients included in the slope limit range of the DP path, and the voiced sound with continuous pitch exists. In the case of no sound, it often does not work well because of lack of smooth pitch continuity even in the case of unvoiced sound.Therefore, after analyzing and synthesizing the voice, adjust the voice phase again. Voice segmentation for so-called segmentation
There is an operational restriction that it is necessary to analyze or synthesize while performing or performing segmentation. Further, the above-mentioned brute force DP method itself has a drawback that the amount of calculation becomes enormous.

The object of the present invention is also to eliminate the above-mentioned drawbacks, to make segmentation fundamentally unnecessary and to significantly reduce operational restrictions and to achieve high flexibility, and also to significantly reduce the amount of calculation.
It is to provide a pitch extraction device based on the DP method.

[Means for solving problems]

The device of the present invention is a pitch extraction device that extracts a pitch period or pitch frequency of a voice through a parameter that expresses the repetitiveness of a voice waveform, and a parameter that is calculated for each preset basic analysis period. A few pitch candidates,
A means for extracting an evaluation value indicating the pitch property of these candidates, a column of the small number of pitch candidates, and a pitch property of these candidates for each section composed of a plurality of preset basic analysis frames From the sequence of evaluation values, the continuity of the sequence of pitch candidates, and means for evaluating the continuity of pitch candidates by limiting the path of each pitch candidate to the previous analysis frame to the period or frequency immediately below and above the pitch candidate. And means for determining the pitch period or the pitch frequency.

〔Example〕

Next, the present invention will be described in detail with reference to the drawings showing an embodiment in which an autocorrelation coefficient is used as a parameter for expressing the repeatability of a speech waveform.

Referring to FIG. 1, the configuration of the first embodiment of the present invention is as follows.
It comprises an autocorrelation coefficient calculator 1, an evaluation value calculator 2, a local maximum value searcher 3, a memory 4, a DP processor 5, and the like.

Upon receiving a voice input, the autocorrelation coefficient calculator 1 samples this at a predetermined sampling frequency, quantizes it with a predetermined number of bits, and then presets a time delay for each 20 mSEC as a basic analysis frame. The autocorrelation coefficient ρi is obtained over the range and is sent to the evaluation calculator 2. The autocorrelation coefficient calculator 1 is obtained in the process of calculating ρi Is output to the evaluation value calculator 2 as electric power P.

The above-mentioned basic analysis frame is set in advance for the time when the speech input can be regarded as almost stationary, and the autocorrelation coefficient ρi is expressed by the following equation (1).

In the equation (1), xj is a sampled value sequence of the quantized speech signal in each basic analysis frame, j is its number, and i is a preset delay time, based on the distribution range of the pitch period to be extracted here. Since the pitch period distribution range is set to 2.5 mSEC to 15 mSEC and the sampling frequency for sampling is 8 KHz in this embodiment, i is determined to be 20 to 120 taps.

N represents the number of reference voice samples, which is the maximum number of i.

Now, the evaluation value calculator 2 inputs the autocorrelation coefficient ρi obtained as described above for each basic analysis frame and calculates the evaluation value gi shown in the following equation (2) as a function of this and the power P. calculate.

gi = (ρi, P) (2) Obtaining the evaluation value gi as the function of the autocorrelation coefficient ρi and the power, the pitch period corresponds to the voiced / unvoiced / silent state via power as faithfully as possible. It is carried out in order to make it extract. As the electric power P used for such a purpose, for example, the formula (1) is used. Or Furthermore, the value may be expressed by any value such as the denominator of Expression (1), and may be set arbitrarily.

The evaluation function gi = (ρi, P) set as a function of the power P and the autocorrelation coefficient ρi is expressed by the following equation (3) or equation (4), for example. Etc. are possible.

As an example of the equation (4), consider the case where the electric power P is represented by the denominator of the equation (1). In this case, gi is the denormalized autocorrelation coefficient. It is also possible to use phi i itself calculated by the self-phase relation calculator 1 as gi, and in this case the evaluation value calculator 2 is the autocorrelation coefficient calculator 1
Can act on your behalf.

The evaluation value gi is supplied to the local maximum value searcher 3, and this evaluation value gi has a large value corresponding to the autocorrelation coefficient ρi, as is clear from the above description. Maximum value searcher 3
Supplies the tap position and the maximum value to the memory 4 while searching for the maximum value of the autocorrelation coefficient sequence for each basic analysis frame thus input via the evaluation value gi.

The memory 4 stores the maximum value of the autocorrelation coefficient sequence for each basic analysis frame input in this way and the data regarding the tap position in a unit of a plurality of predetermined basic analysis frames.

The DP processor 5 reads the maximum value and tap position of the autocorrelation coefficient sequence stored in the memory 4 for each section, sets only the maximum value as a pitch candidate, and the path of each pitch candidate to the previous basic analysis frame is the pitch candidate. The autocorrelation coefficient sequence with tap positions forming the DP path that maximizes this integrated value is desired by limiting the value to the maximum value directly below or above, and using the evaluation value by the function gi shown in equation (2). It is determined and output as the pitch sequence data of.

Now, let us say that the evaluation value in this DP is represented by S (l, i). Here, 1 is the frame number of the basic analysis frame in the section, and i is the search range of the pitch period as described above, which is 20 to 120 taps in this embodiment.

In the DP according to the present invention, the pitch period that provides the optimum DP path that maximizes the integrated value of the DP evaluation values by S (l, i) is solved by solving the pulsation formulas shown by the following formula (5) one after another. Seeking is the basis.

In equation (5), S (l, i) l is the basic analysis frame number
l, the DP evaluation value of the maximum pitch candidate at tap i, p includes the tap position of this maximum value, and is immediately above it, and q
Indicates the tap position of the maximum value of the previous basic analysis frame at the tap position immediately below.

FIG. 2 is a DP path setting explanatory diagram for explaining the contents of the DP path setting in the first embodiment.

It is assumed that the maximum value a at the frame number 1 and at the tap position i is a pitch candidate and its evaluation value is gi. In the immediately preceding basic analysis frame, only the local maxima b and c immediately above and directly below the local maxima a including the tap position i are set as pitch candidates, and the DP evaluation values S (l-1,
p) and S (l−1, q) and gi are used, and a combination of DP evaluation values S (l, i) l that satisfies equation (5) is selected and determined as an optimum value for each maximum value. To be done. The selection of the maximum value of the DP evaluation value is performed for all the basic analysis frames in the section and for all the maximum values, and the final basic analysis frame is reached. This process in normal DP does not take power conditions into consideration, but selects the optimum value for all autocorrelation coefficients within the slope limit range set for pitch changes between each basic analysis frame. Since it is performed, the case of continuous voiced sound that can expect smoothness of pitch continuity is excluded, and if unvoiced sound that often cannot be expected of pitch continuity intervenes, segmentation is assumed as described above. Operation is required, and the amount of calculations for such a brute force DP itself is enormous.

Now, when the selection and determination of the optimum value for all the basic analysis frames in the section and all the local maximum values are processed one after another to reach the final basic analysis frame, the following expression (6) is given. Processing is performed, Smax as the maximum of the DP evaluation values of each maximum value in the final basic analysis frame is determined, and simultaneously with this determination, the integrated value of the DP evaluation values is maximized over all basic analysis frames.
The DP path is automatically determined.

In equation (6), L is the final basic analysis frame number, rt
(T = 1,2, ... μ) is the maximum tap of the final basic analysis frame.

However, the first basic analysis frame does not have the previous basic analysis frame, and therefore the DP evaluation value S (1, i) is all taps 20 to 120.
Denoted by all gi for maxima in the range. Corresponding to this, the DP evaluation value S (2, i) for the second basic analysis frame is expressed by the following equation (7),
In the same manner, one DP path is determined one after another by the method described above.

As described above, only the maximum value of the autocorrelation coefficient sequence in each basic analysis frame is the target of pitch candidates, and the maximum value occupying the tap position immediately above and immediately below the pitch candidate for the path to the previous basic analysis frame of each pitch candidate It is possible to set a stable DP path even if there is no pitch continuity by basically performing segment search by DP that performs pitch search for each segment using the evaluation value by function gi = (ρi, P). This is possible and the amount of calculation can be greatly reduced.

Referring to FIG. 3, the structure of the second embodiment of the present invention is different from the first embodiment shown in FIG. A detailed description of the parts related to the configuration will be omitted.

In the second embodiment, the pitch search is performed through the DP based on the evaluation of a plurality of sections, while the first embodiment described above searches the pitch while performing the DP based on the evaluation value of each section while performing the DP. By searching, the processing delay time and the amount of calculation in the pitch search of voice input are significantly reduced, and the contents are as follows.

The previous section memory 6 is supplied with at least one previous section data for which DP processing is completed from the DP processor 5 through the input / output line 601, and this previous section data is the final basic analysis frame when the previous section is one section. Of the local maximum value data and evaluation values associated with these local maximum values are stored and stored as previous segment data, and when there are two or more previous segments to be stored, the first of two adjacent previous segments The maximum value data and the evaluation value data of the basic analysis frame and the final basic analysis frame of the previous section as a pass partner thereof are also stored together with the data when the previous section is one section according to the number of sections.

FIG. 4 is a DP path setting explanatory diagram for explaining the characteristics of the DP path setting in the second embodiment shown in FIG. 3, and takes the case where the number of previous sections is two as an example. The second embodiment will be described below with reference to FIG.

The previous section data memory 6 is the final basic analysis frame of the previous section, the data regarding the tap position and the evaluation value of each maximum value indicated by the white circle in the frame C shown by the dotted line in FIG. The local maximum value data and the evaluation value data in the first basic analysis frame of the previous section and the final basic analysis frame of the previous previous section included in the basic analysis frame, that is, the frame d indicated by the one-dot chain line are stored. When the number of divisions further increases, the maximum value data and the evaluation value data as shown in the frame d are increased and stored according to the number of adjacent divisions.

By the way, FIG. 4 shows the maximum value indicated by the white circles in the current section, the previous section, and the previous previous section with the frame number, that is, the time and the tap. In the case of FIG. 4, for convenience of explanation, , Each category contains 5 basic analysis frames, and each basic analysis frame has a maximum value of 3
It is modeled as including each one.

Based on the DP method of the present invention described with reference to FIG. 1, the path having the maximum DP evaluation value is selected as the maximum value included in each of these sections, for example, as shown in FIG. In the first embodiment described above, these DPs are
Although this DP was slid over the entire section while being carried out through the DP evaluation value, in the second embodiment, this is carried out based on the evaluation values of the DPs of a plurality of sections. As shown in FIG. 4, the maximum value of the first basic analysis frame of the current section is the path that maximizes the DP evaluation values with the maximum values of the taps immediately below and immediately above the final basic analysis frame of the previous section as shown by the arrow. Is selected, and also between the previous section and the previous section, the same path selection by the DP evaluation value is performed between the first and final basic analysis frames. This second
The DP evaluation value S (1, i) of the first basic analysis frame of each section in the above embodiment is expressed by the following equation (8).

In the equation (8), E represents the final basic analysis frame number of the previous section.

As is clear from the above contents, for example, when considering the DP processing for the same voice input section, the DP processing is performed while sliding the processing for each section over this section of the voice input corresponding to a plurality of sections. In the second embodiment, as compared with the case of the first embodiment to be carried out, the first between adjacent sections is
Therefore, the generation of the delay time due to the processing can be significantly shortened and the amount of calculation can be remarkably reduced only by adding an extremely small amount of time to set the optimum path between the final basic analysis frames.

In each of the first and second embodiments shown in FIGS. 1 and 3, the evaluation value calculator 2 is provided in the preceding stage of the inspection value searcher 3, but the order is reversed. There is no difference in the structure in which the maximum value searcher 3 is provided in the preceding stage of the evaluation value calculator 2, and even if the memory 4 and the previous section data memory 6 are formed as the same structure in FIG. It doesn't matter.

In the above description, the autocorrelation coefficient is used as a parameter that expresses the repetitiveness of the speech waveform, but an AMDF (Average Magnitude-Difference Function) may be used instead of this coefficient. A third embodiment of the present invention using this AMDF is shown in FIG.

Referring to FIG. 5, the third embodiment of the present invention comprises an AMDF calculator 11, an evaluation value calculator 12, a local minimum value searcher 13, a memory 14, a DP processor 15 and a power calculator 16. It

The third embodiment uses AMDF instead of the autocorrelation coefficient used in the first embodiment. In general, the autocorrelation coefficient has a maximum value in the pitch period, and AMDF has a minimum value in the pitch period. Therefore, in the third embodiment, the DP using the minimum value is implemented in place of the DP using the maximum value in the first embodiment, but the third embodiment is basically the same as the first embodiment. Therefore, the difference will be described in particular, and description of common parts will be omitted.

AMDF calculator 11 calculates the AMDF from the input voice using the following equation (9).
Calculate γi.

Regarding AMDF, LR Rabiner, RWSchafer, “Digi
tal Processing of Speech Signal ", Prentice−Hall, In
c, Englewood Cliffs, New Jersey 07632,4,7 “The Short
−Time Average Magnitude−Difference Function ”, pp.
149-150.

The evaluation value calculator 12 receives AMDFγi as input and the power calculator
The evaluation value φi shown in the following equation (10) is calculated as a function of the electric power P supplied from 16.

φi = (γi, P) (10) Note that φi is an evaluation value that generally decreases monotonically with respect to P. The local minimum value searcher 13 searches for local minimum values of φi, for example, only three, and outputs the delay time corresponding to these values to the memory 14. The memory 14 stores these data in division units and
Output to the processor 15.

The DP processor 15 reads the data in the memory 14 and searches for the pitch in the same manner as the DP processor 5 in the first embodiment. However, the DP path with the smallest DP evaluation value is selected.

The power calculator 16 calculates power from the input voice and outputs it to the evaluation value calculator 12 as described above.

〔The invention's effect〕

As described above, according to the present invention, the pitch extraction based on the DP method, which basically eliminates the execution of segmentation, significantly improves the operability, significantly improves the time delay due to the processing, and significantly reduces the required calculation amount, is also provided. There is an effect that the device can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a first embodiment of the present invention, and FIG. 2 is a DP path setting explanatory diagram for explaining contents of a DP path setting in the first embodiment of FIG. FIG. 3 is a block diagram showing the configuration of the second embodiment of the present invention, and FIG. 4 is a DP path setting explanatory diagram for explaining the characteristics of the DP path setting in the second embodiment shown in FIG. FIG. 5 is a block diagram showing the configuration of the third embodiment of the present invention. 1 ... Autocorrelation coefficient calculator, 2 ... Evaluation value calculator, 3 ...
... maximum value searcher, 4 ... memory, 5 ... DP processor,
6 ... Previous section data memory, 11 ... AMDF calculator, 12 ...
Evaluation value calculator, 13 ... Minimum value calculator, 14 ... Memory, 15
…… DP processor, 16 …… Power calculator.

Claims (8)

[Claims] 1. A pitch extraction device for extracting a pitch period or pitch frequency of a voice through a parameter expressing the repetitiveness of a voice waveform, wherein the parameter is calculated for each preset basic analysis period. A unit for extracting a small number of pitch candidates and an evaluation value indicating the pitch property of these candidates, a column of the small number of pitch candidates, for each section composed of a plurality of preset basic analysis frames, and these From the sequence of evaluation values indicating the pitch characteristics of the candidates, the continuity of the sequence of pitch candidates and the succession of pitch candidates by limiting the path of each pitch candidate to the previous analysis frame to the period or frequency immediately below and above the pitch candidate. And a means for determining a pitch period or a pitch frequency by using a means for evaluating the property. 2. The pitch extracting apparatus according to claim 1, wherein the pitch search is performed while obtaining the evaluation value for a plurality of the sections. 3. A means for evaluating continuity of a pitch candidate by limiting a path of each pitch candidate to a previous analysis frame to a period or a frequency immediately below and above the pitch candidate, respectively. The pitch extraction device according to the item 1). 4. A pitch extracting apparatus according to claim 1, wherein an autocorrelation coefficient is used as a parameter expressing the repetitiveness of a speech waveform. 5. A maximum value searching means for searching a maximum value of an autocorrelation coefficient sequence calculated for each preset basic analysis frame and storing the tap position and the maximum value, and only the maximum value as a pitch candidate. The path of each pitch candidate to the previous analysis frame is limited to the maximum values of the tap positions directly below and above the pitch candidate, and the pitch search is executed by the dynamic programming method for each of the preset plurality of sections of the basic analysis frame. The pitch extracting apparatus according to claim (4), further comprising a dynamic programming execution means. 6. The pitch extracting apparatus according to claim 5, further comprising means for setting a value obtained by weighting the maximum value with power information as an evaluation value. 7. An AMDF (Average Magnitude-Difference functio) is used as a parameter expressing the repetitiveness of a speech waveform.
n) is used. Claim (1) characterized in that
The pitch extraction device according to the item. 8. A minimum value search means for searching a minimum value of AMDF calculated for each preset basic analysis frame and storing the tap position and the minimum value, and each of the pitch candidates having only the minimum value as a pitch candidate. Dynamic planning in which the path to the previous analysis frame of the basic analysis frame is limited to the minimum values of the tap positions immediately below and above the pitch candidate and the pitch search is executed by the dynamic programming method for each of the preset plurality of sections of the basic analysis frame. A pitch extracting apparatus according to claim (7), further comprising a law executing means.