JPH09258762A - Pitch extracting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To extract the pitch of voice waveform with a small quantity of operation by restricting an interval to be a prospect in waveform at the time of computing a square error, so as to become an interval starting from the value approximate to the head value of waveform data to be examined. SOLUTION: A prospective value computing part 11 obtains the head value αof waveform data to extract a pitch out of digital voice waveform data stored in a data buffer 2, and computes the prospective value α± for pitch extraction on the basis of the head value α. A pitch prospect extracting part 12 searches out each position of the waveform data with the same value as the prospective value α±5 from the buffer 2. The interval length from the head value α to each position becomes the interval length of each pitch prospect. A square error computing part 13 computes the square error of each pitch prospect and outputs the pitch prospect with the square error having the shortest inverval length, as an extracted pitch.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch extracting device for extracting a pitch of a voice waveform for analyzing and compressing voice data.

[0002]

2. Description of the Related Art A voice waveform has a basic vibration period,
This is generally called pitch. This pitch is frequently extracted when voice data is analyzed or compressed, and various extraction methods have been proposed. As the simplest algorithm among those pitch extraction methods, there is a method of calculating the squared error of two predetermined sections in the waveform and using the section length having the smallest double error as the pitch.

FIGS. 6A and 6B are views for explaining a pitch extraction method based on a square error in the conventional pitch extraction method of this type. FIG. 6A shows the case where the waveform is cut out in the section of the section length m, and in this case, the double error between the two waveforms becomes large.

Further, FIG. 6B shows a case where a waveform is cut out in a section having a section length n, and in this case, the double error between the two waveforms becomes small. As described above, in the case of the waveform shown in FIG. 6, the square error is smaller in the case of the section length n than in the case of the section length m, and it can be said that it is closer to the pitch length.

[0005]

As a problem of the pitch extraction method by calculating the squared error as described above, the longer the pitch length to be investigated (target section length) is,
There is a drawback that the amount of calculation for calculating the squared error increases. However, if the target section length is limited to be short, the pitch of the waveform having a low frequency cannot be extracted.

Further, since it is not known which section length corresponds to the pitch, it is necessary to obtain the square error for all the section lengths within the range that can be considered as the pitch. For this reason, a large amount of calculation is required, which becomes a problem in an environment where only a small amount of power is available in the CPU that performs those calculations. It is an object of the present invention to provide a pitch extraction device that can extract the pitch of a voice waveform with a small amount of calculation.

[0007]

In order to solve the above problems and achieve the object, the pitch extracting apparatus of the present invention is configured as follows. (1) The pitch extraction device of the present invention is obtained by accumulating means for accumulating the waveform data, obtaining means for obtaining the value of the first data in the waveform data accumulated by the accumulating means, and the obtaining means. Searching means for searching the waveform data for a plurality of data having a value close to the value, and the length of a section formed by the position of the leading data and the position of each data searched by the searching means in the waveform data Is defined as the section length, and the predetermined section length is formed from the first waveform forming a predetermined section length from the position of the leading data and the position of the data searched by the searching means corresponding to the predetermined section length. A detection means for detecting a difference from the second waveform,
The section length with the smallest difference detected by the detection means is extracted as the pitch. (2) The pitch extracting device of the present invention is the device according to (1) above, and further comprises a calculating means for adding or subtracting a predetermined value to or from the value acquired by the acquiring means, and is calculated by the calculating means. The search means is configured to search for a plurality of data having a value close to the value. (3) The pitch extraction device of the present invention is the device according to (2) above, and further comprises a second acquisition means for acquiring a dynamic range of the waveform data accumulated by the accumulation means, and a value of the dynamic range. And a second calculation means for dividing by the preset value, and the value calculated by the second calculation means is set as the predetermined value. (4) The pitch extracting device of the present invention has the above (1) to (3).
The apparatus according to any one of claims 1 to 5, wherein the detection means is configured to detect a squared error between the first waveform and the second waveform.

As a result of taking the above-mentioned means, the following actions are brought about respectively. (1) According to the pitch extracting apparatus of the present invention, the length of the section formed by the positions of a plurality of data having values close to the acquired value and the position of the head data is set as the section length, and A difference between the first waveform having a predetermined section length from the data position and the second waveform having the predetermined section length from the data position corresponding to the predetermined section length is detected, and the difference is the smallest. Since the section length is extracted as the pitch, it is possible to easily narrow down the section length as a pitch candidate, and it is possible to extract the pitch from the narrowed down section length with an average small calculation amount. Therefore, even in an environment in which a small amount of power is available in the CPU that performs those calculations, for example, the pitch of a waveform having a low frequency can be extracted without any problem. (2) According to the pitch extracting apparatus of the present invention, a predetermined value is added to or subtracted from the acquired value, and a plurality of data having a value close to the calculated value is searched for. Further appropriate narrowing down of the section length can be performed in consideration of the fluctuations that occur and the deviations that occur when digitizing each value of the waveform data. (3) According to the pitch extraction device of the present invention, the value obtained by dividing the value of the dynamic range of the accumulated waveform data by the preset value is set as the predetermined value, and therefore the predetermined value is appropriately set according to the waveform. It is possible to narrow down the section length more appropriately. (4) According to the pitch extracting apparatus of the present invention, since the square error between the first waveform and the second waveform is detected,
A general pitch extraction method can be applied to the pitch extraction device.

[0009]

1 is a block diagram showing the configuration of a pitch extracting device according to an embodiment of the present invention. FIG.
1 is a control unit, and the control unit 1 includes a candidate value calculation unit 11, a pitch candidate extraction unit 12, and a squared error calculation unit 1.
Consists of three. The candidate value calculation unit 11, the pitch candidate extraction unit 12, and the squared error calculation unit 13 are each connected to the data buffer 2. A pitch candidate extraction unit 12 is connected to the candidate value calculation unit 11, and a squared error calculation unit 13 is connected to the pitch candidate extraction unit 12. Further, a ROM 3 and a RAM 4 are connected to the control unit 1.

An input waveform (digital audio signal) is temporarily stored in the data buffer 2 as waveform data for pitch extraction calculation. The control unit 1 controls the operation of the pitch extracting device. The candidate value calculation unit 11 in the control unit 1 reads the value (for example, α) of the first data of the waveform data stored in the data buffer 2, and based on the value, the candidate value for pitch extraction ( For example, a value in the range of α ± 5) is calculated. The necessity of calculating this candidate value will be described later.

The pitch candidate extraction unit 12 searches the data buffer 2 for a plurality of sample data having the same value as the candidate value calculated by the candidate value calculation unit 11. And
The length of the section from the beginning of the data buffer 2 to each of the sample data is a pitch candidate. The square error calculating unit 13 calculates the square error of the waveform in each pitch candidate input from the pitch candidate extracting unit 12, extracts the pitch candidate having the smallest double error, and outputs it as the pitch.

The ROM 3 stores the operation control program of the pitch extracting device, and the RAM 4 stores the operation control program.
As will be described later, the value of the waveform data for each pitch candidate, which is required when the squared error calculator 13 calculates the squared error, is stored.

FIG. 2 is a diagram showing an example of a voice waveform for performing pitch extraction in the pitch extraction device. In the pitch extracting device, as described above, the data buffer 2
Pitch extraction is performed based on the digital audio signal stored in FIG. 2, but for convenience sake, an analog audio waveform is shown and described in FIG. 2 and FIG. 4 described later.

Since the pitch is the fundamental vibration period as described above, the value (P) of the beginning data of the voice waveform data to be examined and the value (Q) of the voice waveform data at the start point of the next pitch 1 are: It is considered that the value is usually approximate. Therefore, in the pitch extracting apparatus, when calculating the squared error, the section that becomes a candidate in the waveform is narrowed down to a section that starts from a value close to the value of the first data of the waveform data to be examined. This makes it possible to reduce the average amount of calculation required for pitch extraction as compared with the conventional extraction method.

FIG. 3 is a flow chart showing a processing procedure in the pitch extracting apparatus. Hereinafter, the processing procedure of the pitch extracting apparatus as described above will be described with reference to FIG.
First, in step S1, the candidate value calculation unit 11 acquires the leading value of the waveform data from which the pitch is to be extracted, of the digital voice waveform data stored in the data buffer 2.

Next, in step S2, the leading value needs to have a certain width as will be described later, so the candidate value calculation unit 11 determines a candidate value for pitch extraction based on the leading value. calculate. For example, when the leading value is α and the width is 5, values in the range of α ± 5 are candidate values.

Subsequently, in step S3, the pitch candidate extraction unit 12 stores the respective positions of a plurality of data in the waveform data having the same value as the candidate value (α ± 5) in the voice waveform stored in the data buffer 2. Search from data. Thereby, the length of the section from the first value to each position is
Each becomes a pitch candidate section length. And step S4
Then, the square error calculation unit 13 calculates the square error of each pitch candidate, and in step S5, the pitch candidate having the section length with the smallest double error among the calculated double errors is extracted. Output as pitch.

FIG. 4 is a diagram showing a voice waveform for explaining the narrowing down of the above-mentioned candidate section length. Less than,
It is assumed that a section starting from data having a value close to the value of the data O at the head of the voice waveform data shown in FIG. 4 is narrowed down. In the example of FIG. 4, the data having a value approximate to the value of the data O at the beginning of the waveform data is A to E.

Since it is considered that there is a pitch in the section length from the position of the head data O to any position of the data A to E, the section length from O to A is a and the section length from O to B is a. The length is set to b ... The same applies hereinafter, and the square error of the waveform for two continuous section lengths is calculated.

That is, each squared error of the waveforms of the two section lengths a and a ', b and b', ... E and e'shown in FIG. 4 is calculated. Then, in this example, the square error of the waveforms of the section lengths d and d ′ is the smallest, so that the result is that d is the pitch to be extracted. When the square error calculator 13 calculates the square error, the following data is stored in the RAM 4.

FIG. 5 is a memory map showing a storage state of data for calculating a squared error in the RAM 4. The RAM 4 stores the first value (START POINT) of the waveform first extracted by the candidate value calculation unit 11, and hereinafter stores a plurality of data values in each section length a to e.

For example, the data value in the section length a is A0
.About.A10, and a total of 11 data values are stored, each including A and five data before and after it. Similarly, for the section lengths b to e, 11 corresponding data values are stored. In this way, the double error of each section length a to e is calculated based on the value of each data stored in the RAM 4.

When the pitch is actually extracted, the pitch length fluctuates with the passage of time, or each value of the waveform data shifts when digitized by sampling or the like. Therefore, it is necessary not only to narrow down the candidate sections based on the same value as the leading data of the waveform data to be examined, but also to allow the values to have some width. The required width varies depending on the range of the waveform, the number of quantization bits when digitized, and the like, so it is necessary to appropriately determine it based on the input voice waveform data.

In the present embodiment, the eigenvalue (α) ± 5 range is used to calculate the candidate value. However, for example, the dynamic range of the waveform is acquired, and the value obtained by dividing the value by a predetermined value is obtained. It is also possible to narrow down by appropriately changing the candidate value by adding / subtracting (that is, several% of the dynamic range) to the value of the leading data.

The present invention is not limited to the above-described embodiments, but can be modified and implemented as appropriate without departing from the scope of the invention. (Summary of Embodiment) The configuration, operation and effect shown in the embodiment are summarized as follows. [1] The pitch extraction device shown in the embodiment has a storage unit (2) for storing waveform data, and a storage unit (2).
Acquisition means (11) for acquiring the value of the first data in the waveform data accumulated in step (1), and searching the waveform data for a plurality of data having values close to the values acquired by the acquisition means (11). Searching means (12), the position of the top data in the waveform data, and the searching means (1
The length of the section formed by the position of each data searched in 2) is set as the section length, and the first waveform forming a predetermined section length from the position of the leading data and the above-mentioned predetermined section length corresponding to the first waveform. A detection unit (13) for detecting a difference from the position of the data searched by the search unit (12) and the second waveform having the predetermined section length, and the detection unit (13) detects the difference. The interval length with the smallest difference is extracted as the pitch.

As described above, in the pitch extracting apparatus, the length of the section formed by the position of the plurality of data having a value close to the acquired value and the position of the head data is defined as the section length, and A difference between the first waveform having a predetermined section length from the data position and the second waveform having the predetermined section length from the data position corresponding to the predetermined section length is detected, and the difference is the smallest. Since the section length is extracted as the pitch, it is possible to easily narrow down the section length as a pitch candidate, and it is possible to extract the pitch from the narrowed down section length with an average small calculation amount. Therefore, even in an environment in which a small amount of power is available in the CPU that performs those calculations, for example, the pitch of a waveform having a low frequency can be extracted without any problem. [2] The pitch extraction device shown in the embodiment is the device according to the above [1], and the acquisition means (1
1) is provided with a calculating means (11) for adding or subtracting a predetermined value to the value, and the searching means (12) searches for a plurality of data having a value approximate to the value calculated by the calculating means (11). It is configured to explore.

As described above, in the above pitch extracting device, a predetermined value is added to or subtracted from the acquired value to search for a plurality of data having a value close to the calculated value, so that the pitch length time elapses. Further appropriate narrowing down of the section length can be performed in consideration of the fluctuations that occur and the deviations that occur when digitizing each value of the waveform data. [3] The pitch extracting device shown in the embodiment is the device according to [2] above, and the accumulating means (2) is provided.
Second acquisition means (11) for acquiring the dynamic range of the waveform data accumulated in 1. and second calculation means (1) for dividing the value of the dynamic range by a preset value.
1) and is configured to set the value calculated by the second calculating means (11) as the predetermined value.

As described above, in the pitch extracting apparatus, the value obtained by dividing the value of the dynamic range of the accumulated waveform data by the preset value is used as the predetermined value, and therefore the predetermined value is appropriately set according to the waveform. It is possible to narrow down the section length more appropriately. [4] The pitch extraction device shown in the embodiment is the device according to any one of the above [1] to [3], and the detection means (13) has the first waveform and the first waveform. Two
Is configured to detect the squared error formed by the waveform of

As described above, since the above-mentioned pitch extracting apparatus detects the square error between the first waveform and the second waveform, a general pitch extracting method can be applied to the pitch extracting apparatus. it can.

[0030]

As described above, according to the present invention, it is possible to provide a pitch extracting device which can extract the pitch of a voice waveform with a small amount of calculation.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a pitch extraction device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a voice waveform for performing pitch extraction in the pitch extraction device according to the embodiment of the present invention.

FIG. 3 is a flowchart showing a processing procedure in the pitch extraction device according to the embodiment of the present invention.

FIG. 4 is a diagram showing a voice waveform for explaining narrowing down of a section length that is a candidate according to the embodiment of the present invention.

FIG. 5 is a memory map showing a storage state of data for calculating a squared error in RAM 4 according to the embodiment of the present invention.

FIG. 6 is a diagram for explaining a pitch extraction method based on a square error in the pitch extraction method according to the conventional example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Control part 11 ... Candidate value calculation part 12 ... Pitch candidate extraction part 13 ... Double error calculation part 2 ... Data buffer 3 ... ROM 4 ... RAM

Claims (4)

[Claims] 1. A storage means for storing waveform data, an acquisition means for acquiring the value of the first data in the waveform data stored by the storage means, and a value approximating the value acquired by the acquisition means. Searching means for searching a plurality of data from the waveform data, and the length of a section formed by the position of the top data and the position of each data searched by the searching means in the waveform data, respectively, as a section length, A first waveform having a predetermined section length from the position of the leading data and a second waveform having the predetermined section length from the position of the data searched by the searching means corresponding to the predetermined section length. A pitch extracting device, comprising: a detecting unit that detects a difference, wherein a section length having the smallest difference detected by the detecting unit is extracted as a pitch. 2. A search means for searching for a plurality of data having a value approximate to the value calculated by the calculation means, comprising calculation means for adding or subtracting a predetermined value to the value acquired by the acquisition means. The pitch extracting device according to claim 1, wherein: 3. A second acquisition means for acquiring the dynamic range of the waveform data accumulated by the accumulation means, and a second arithmetic means for dividing the value of the dynamic range by a preset value. The pitch extraction device according to claim 2, wherein the value calculated by the second calculation means is the predetermined value. 4. The pitch extracting apparatus according to claim 1, wherein the detecting means detects a squared error formed between the first waveform and the second waveform.