JPH0743600B2 - Pitch extractor - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pitch extraction apparatus, and more particularly, to a pitch period (frequency) search section divided by a geometric series to maximize the repeatability of speech for each divided section. Select things based on DP method,
A pitch extracting device for determining.

[Conventional technology]

Autocorrelation method using autocorrelation coefficient, or AMDF (Av
erage Multitude Difference Function) method, etc., to maximize the repetitiveness of speech (autocorrelation method) or minimum (AMD)
The pitch extraction that extracts the pitch period while searching the maximum value for each predetermined search section by the DP method, using the evaluation value obtained based on the predetermined evaluation function for the extreme value of F) and its tap position as the evaluation scale. The device is known to be capable of greatly reducing pitch extraction errors, and this is described in detail, for example, in Japanese Patent Application No. 61-212691, "Pitch Extractor".

[Problems to be solved by the invention]

The above-described conventional pitch extracting apparatus described above can significantly reduce the pitch extraction error by the DP method, but on the other hand, it has the drawback that the amount of calculation is unavoidable. The reason for this is that the DP method is based on the essential feature of the DP method, in which all the maximum repeatability of speech is applied to the DP processing as pitch period candidates.

An object of the present invention is to eliminate the above-mentioned drawbacks, and to provide a pitch extraction device capable of remarkably reducing the amount of calculation by dividing the search interval of the pitch period in geometric progression and performing pitch extraction by the DP method. It is in.

[Means for solving problems]

The apparatus of the present invention comprises a coefficient extracting means for extracting a coefficient expressing the repeatability of speech, and a maximum or minimum extreme value which differs depending on the type of the coefficient of the coefficient string expressing the repeatability obtained by the coefficient extracting means. And a pitch characteristic evaluation means for performing pitch characteristic search of the voice by DP (Dynamic Programming) processing while obtaining an evaluation value based on a predetermined evaluation function for the tap position and a pitch to be searched by the pitch characteristic evaluation means. The search section of the sex is divided into divided sections by a geometric series of preset ratios, and only the maximum one maximum value or the minimum one minimum value is set for each divided section. DP
And a pitch period determining means for determining a pitch period while executing the processing and selecting the one having the maximum repeatability.

〔Example〕

 Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

The configuration of the embodiment shown in FIG. 1 has an A / D converter 1 which quantizes an input voice and outputs it for each predetermined analysis frame, and a self-correlation coefficient sequence as a coefficient sequence expressing the repeatability of the input voice. Stores the section information of the correlation calculator 2 and the maximum value searcher 3 for searching the maximum value for each search section and the tap position for the autocorrelation coefficient string output from the autocorrelation calculator 2 and the maximum section value searcher 3 The divided section data memory 4, the maximum value of each search section output from the section maximum value searcher 3 and the tap position thereof are stored in the memory 5, and the maximum value of each search section read from the memory 5 and the data regarding the tap position are stored. On the other hand, it comprises a DP processor 6 which performs selection by the DP method and extracts the pitch period.

Next, the operation of the pitch extracting device of this embodiment will be described.

The embodiment shown in FIG. 1 exemplifies a case where an autocorrelation coefficient is used as a coefficient expressing the repeatability of input speech.

The A / D converter 1 samples the input voice at a sampling frequency of 8 KHz and then quantizes it with a predetermined number of bits,
Output in 20 msec basic analysis frame units.

The autocorrelation calculator 2 calculates the autocorrelation coefficient ρ over a predetermined time delay range for each basic analysis frame of the quantized speech input.
_i is determined by the following equation (1).

In the equation (1), x _j is a sample string of the quantized speech signal for each basic analysis frame, j is its serial number, i is a predetermined delay time set in advance, and the pitch period distribution range is 2 msec to 15 msec. Since i is set and the sampling frequency is 8 KHz, i is 16 to 121 taps. The maximum tap obtained from 8 KHz 15 msec is 120 taps, but it is set to 121 by taking into account the geometrical setting condition of the divided section described later.

The autocorrelation calculator 2 determines the number of taps for each basic analysis frame.
An autocorrelation coefficient sequence according to the equation (1) is calculated within the delay time range of 16 to 121, and this is supplied to the maximum section search unit 3.

The section maximum value searcher 3 decomposes the delay time range of the number of taps 16 to 121 into preset decomposition sections, and has the highest level among the maximum values of the autocorrelation coefficient for each divided section, that is, The maximum value is retrieved, and the maximum value and the tap position information for each divided section are sent to the memory 5 and stored therein.

In the present invention, the above-mentioned division intervals of the division areas are set so that the mutual interval ratio is geometrically set, and in the present embodiment, the division ratio is set to 1.5, and the division interval by this is 16
〜24,24〜36,36〜54,54〜81,81〜121
In this case, the reason for setting the geometric series of 1.5 times for each tap section is as follows.

That is, the main cause of the pitch extraction error is that a double pitch error, that is, a double pitch of a correct pitch cycle is mistaken as a pitch cycle. Therefore, it may be assumed that there is no autocorrelation larger than the autocorrelation of the basic pitch in a section that is less than twice the correct pitch period. In the present embodiment, focusing on this feature, instead of performing DP processing for all maximum values in the delay time range of 16 to 121 taps by the DP method, the search section is equalized so that the section length is less than twice the pitch period. Divide by a geometric series, search for the maximum value of the maximum value for each divided interval, and target only these maximum values
This is DP processing.

The minimum and maximum tap ratios of the divided sections are set to 1.5 in consideration of some fluctuation of the pitch period in addition to the consideration of the double pitch error described above.

Since the number of section divisions determined in this way is much smaller than the number of maximum values, the amount of calculation required for DP processing that is performed only on the maximum value set for each division section is also significant. Can be reduced.

Data for each DP processing execution section is output from the section maximum value searcher 3 and stored in the memory 5.

The data stored in the memory 5 is read by the DP processor 6 and subjected to DP processing.

The DP processor 6 reads the maximum value of the autocorrelation coefficient and the tap position for each divided section, and sets these as pitch candidates,
DP that maximizes the integrated value of the evaluation values by the predetermined evaluation function g _i
An autocorrelation coefficient string having a tap position that provides a path is determined as desired pitch string data, and a pitch period based on this pitch string data is output.

As the evaluation function g _i for determining the DP path, the one shown in the following expression (2) is used in this embodiment.

g _i = f (ρ _i , P) (2) The meaning of the equation (2) is the autocorrelation coefficient ρ _i
And the short-term voice power P indicated by the autocorrelation coefficient at zero delay time, or Alternatively, the short-time voice power P represented by the denominator or the like of the equation (1)
G _i , which is set as a function of and, is used as the evaluation function, and is actually expressed by the following equation (3) or (4).
Use g _i1 or g _i2 .

g _i1 = ρ _i · logP (3) g _i2 = ρ _i · P (4) As described above, the evaluation function is used as the product of ρ _i and P when it is voice / unvoiced. Since the level of the short-time voice power corresponds, the voiced / unvoiced information closely related to the pitch period is reflected via this short-time voice power to obtain the evaluation value necessary for the DP path setting. Generally, such use of the short-time voice power P is not an essential condition.

In the above, the case where an autocorrelation coefficient is used as a coefficient that expresses the repeatability of speech has been taken as an example, but instead of this, the repeatability is expressed using a coefficient of another format, for example, data obtained by the AMDF method. However, it is obvious that the same operation can be performed.

〔The invention's effect〕

As described above, according to the present invention, when the pitch period is obtained by the DP method using the coefficient expressing the repeatability of speech, the maximum value of each of the intervals obtained by dividing the search interval of the pitch period by the geometric series is used. By significantly reducing the DP processing data by means of executing the DP processing only for the maximum or the minimum minimum value and selecting the one with the maximum repeatability of the voice, the DP processing calculation amount can be reduced. There is an effect that it is possible to realize a pitch extraction device that can be significantly reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. 1 ... A / D converter, 2 ... Autocorrelation calculator, 3 ...
... section maximum value searcher, 4 ... divided section data memory, 5
...... Memory, 6 ... DP processor.

Claims (1)

[Claims] 1. A coefficient extracting means for extracting a coefficient expressing the repeatability of speech, and a maximum or minimum extreme value which differs depending on the kind of the coefficient of a coefficient string expressing the repeatability obtained by the coefficient extracting means, and Pitch property evaluation means for performing pitch property search of the voice by DP (Dynamic Programming) processing while obtaining an evaluation value based on a predetermined evaluation function for the tap position, and pitch property to be searched by the pitch property evaluation means. The search section is defined as a division section that is divided by a geometric series of a preset ratio, and only the maximum of one maximum value or the minimum of one minimum value is set for each DP. A pitch period determining means for determining a pitch period while executing a process and selecting the maximum repeatability. Chi extractor.