Feature quantity extraction method, determination method, and feature quantity extraction device

Abstract

To provide a technique for suitably extracting feature quantities contained in audio signals.SOLUTION: A method comprises the steps of: receiving a voice signal that may include a component whose frequency periodically changes within a certain range; and inputting the received voice signal into a filter bank 12 to extract a feature quantity of the voice signal. When one of a plurality of bandpass filters in the filter bank 12 is denoted as I-th bandpass filter and a bandpass filter adjacent to the I-th bandpass filter in the frequency domain is denoted as (I+1)th bandpass filter, a width Di of the I-th bandpass filter, an interval Li(i+1) between a center frequency fi of the I-th bandpass filter and a center frequency fi+1 of the (I+1)th bandpass filter, and a peak value Pi of the I-th bandpass filter are set based on a width di of the possible frequency changes at the center frequency fi of the I-th bandpass filter.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a feature extraction technique, and more particularly to a feature extraction method for extracting features contained in an audio signal, a determination method, and a feature extraction device. [Background technology]

[0002] When a blade of a wind turbine has a scratch or a hole, a phenomenon occurs in which a high-pitched sound called a whistling noise is generated. To detect the whistling noise, for example, a time-dependent transition of a frequency component with high sharpness is detected as a Doppler shift (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2010-281279 A Summary of the Invention [Problem to be solved by the invention]

[0004] When there is external noise or a difference in volume due to rotation, the whistling sound may not be continuous but may only be acquired intermittently. In such a situation, the Doppler shift cannot be detected, and therefore the whistling sound cannot be detected.

[0005] The present invention has been made in view of the above circumstances, and has an object to provide a technique for suitably extracting a feature contained in an audio signal. [Means for solving the problem]

[0006] In order to solve the above problem, a feature extraction method according to an embodiment of the present invention includes the steps of: receiving an audio signal that may include a component whose frequency changes periodically within a certain range; and inputting the received audio signal into a filter bank that is an array of a plurality of band-pass filters to extract features of the audio signal. If one of the band-pass filters in the filter bank is designated as the I-th band-pass filter, and a band-pass filter adjacent to the I-th band-pass filter in the frequency domain is designated as the I+1-th band-pass filter, the width Di of the I-th band-pass filter, the interval Li(i+1) between the center frequency fi of the I-th band-pass filter and the center frequency fi+1 of the I+1-th band-pass filter, and the peak value Pi of the I-th band-pass filter are set based on the width di of the frequency change that can be obtained at the center frequency fi of the I-th band-pass filter.

[0007] Another aspect of the present invention is a feature extraction device. The device includes an input unit that receives an audio signal that may include a component whose frequency changes periodically within a certain range, and a filter bank that is an array of a plurality of band-pass filters that receives the received audio signal and extracts a feature of the audio signal. In the filter bank, when one of the plurality of band-pass filters is designated as an I-th band-pass filter and a band-pass filter adjacent to the I-th band-pass filter in the frequency domain is designated as an I+1-th band-pass filter, the width Di of the I-th band-pass filter, the interval Li(i+1) between the center frequency fi of the I-th band-pass filter and the center frequency fi+1 of the I+1-th band-pass filter, and the peak value Pi of the I-th band-pass filter are set based on the width di of the frequency change that can be obtained at the center frequency fi of the I-th band-pass filter.

[0008] Any combination of the above components, and any transformation of the present invention into a method, device, system, recording medium, computer program, etc., are also effective as aspects of the present invention. Effect of the Invention

[0009] According to the present invention, it is possible to suitably extract a feature amount contained in a voice signal. [Brief description of the drawings]

[0010] [Figure 1] 1 is a diagram showing a configuration of a feature extraction and determination device according to an embodiment. [Diagram 2] FIG. 2 is a diagram showing the characteristics of the bandpass filter of FIG. [Diagram 3] 3(a) and 3(b) are diagrams showing feature quantities output from each band-pass filter in FIG. [Figure 4] FIG. 2 is a diagram showing the configuration of a wind power generator to be determined; [Diagram 5] 5(a)-(c) are diagrams showing an overview of the processing for the wind power generator in FIG. [Figure 6] FIG. 2 is a diagram showing an overview of a process performed by a determination unit in FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Before describing the present invention in detail, an overview will be given first. An embodiment of the present invention relates to a feature extraction and determination device that extracts features of an audio signal and determines whether the audio signal is normal or not based on the features. As described above, when extracting features based on the Doppler shift, it is difficult to extract the features in a situation where the Doppler shift cannot be detected. One of the feature extraction methods that does not use the Doppler shift is the Mel filter bank. The Mel filter bank arranges multiple bandpass filters in parallel. However, since each bandpass filter in the Mel filter bank is weighted according to the human pitch perception scale, it is not suitable for analyzing frequency changes caused by physical phenomena such as the Doppler effect.

[0012] In this embodiment, a filter bank in which a plurality of band-pass filters are arranged in parallel is used, similar to a Mel filter bank. However, in the filter bank according to this embodiment, the parameters of each band-pass filter are set based on the range of change under the condition that the speech signal to be extracted changes within a certain range.

[0013] 1 shows the configuration of a feature extraction / determination device 100. The feature extraction / determination device 100 includes an input section 10, a filter bank 12, and a determination section 14. The filter bank 12 includes a first band-pass filter 20a, a second band-pass filter 20b, a third band-pass filter 20c, . . . , an N-th band-pass filter 20n, which are collectively referred to as band-pass filters 20, and a synthesis section 22.

[0014] The feature extraction and determination device 100 is realized by any information processing device, such as a general-purpose server (including a distributed server) or a PC (Personal Computer). Alternatively, for example, the feature extraction and determination device 100 may be realized by a smartphone or a tablet terminal, or a semiconductor chip such as a SoC (System on Chip). For example, when the feature extraction and determination device 100 is realized by a PC, the PC may be a dedicated computer designed to realize the feature extraction and determination device 100, or may be a general-purpose computer.

[0015] The input unit 10 receives an audio signal via a microphone (not shown). The audio signal may contain a component whose frequency changes periodically within a certain range due to, for example, a physical phenomenon. The audio signal is received for a period of, for example, about one hour. It is desirable that the period for receiving the audio signal is relatively long. However, if the period for receiving the audio signal is short, it is desirable that the length of the audio signal is equal to or longer than the length of the period in which the frequency changes within a certain range due to a physical phenomenon, or equal to or longer than an integer multiple of the period.

[0016] The filter bank 12 is composed of an array of a plurality of band pass filters 20. The band pass filters 20 input the received audio signal to each of the plurality of band pass filters 20. The first band pass filter 20a, the second band pass filter 20b, the third band pass filter 20c, ..., the Nth band pass filter 20n are arranged in order from low to high frequency. The filter bank 12 synthesizes the outputs from each of the plurality of band pass filters 20 by the synthesis unit 22. Through such processing, the filter bank 12 extracts the feature amount of the audio signal. Here, FIG. 2 is also used to explain in detail the characteristics set in each band pass filter 20.

[0017] FIG. 2 shows the characteristics of the bandpass filter 20. The horizontal axis indicates frequency, and the vertical axis indicates weight. Here, one of the bandpass filters 20 in the bandpass filter 20 is indicated as the I-th bandpass filter 20i, and the filter characteristics of the I-th bandpass filter 20i are indicated as the I-th bandpass filter characteristics 120i. In addition, the I+1-th bandpass filter 20i+1, the I+2-th bandpass filter 20i+2, and the I+3-th bandpass filter 20i+3 are arranged in order from the I-th bandpass filter 20i in the direction of increasing frequency. The I+1-th bandpass filter 20i+1 is adjacent to the I-th bandpass filter 20i in the frequency domain. The filter characteristics of the I+1-th bandpass filter 20i+1 are indicated as the I+1-th bandpass filter characteristics 120i+1, and the filter characteristics of the I+2-th bandpass filter 20i+2 are indicated as the I+2-th bandpass filter characteristics 120i+2. The filter characteristic of the (I+3)th bandpass filter 20i+3 is indicated as the (I+3)th bandpass filter characteristic 120i+3.

[0018] The center frequency of the I-th bandpass filter 20i is indicated as "fi", the width of the I-th bandpass filter 20i is indicated as "Di", and the interval between the center frequency fi of the I-th bandpass filter 20i and the center frequency fi+1 of the I+1-th bandpass filter 20i+1 is indicated as "Li(i+1)". The peak value of the I-th bandpass filter 20i is indicated as "Pi". Pi corresponds to the weight in fi. The integral value of the I+1-th bandpass filter characteristic 120i+1 is indicated as Di×Pi / 2. This integral value is the same for multiple bandpass filters 20.

[0019] The center frequency fi, the width Di, the interval Li(i+1), and the peak value Pi are summarized as the I-th band-pass filter characteristic 120i described above. The following describes the setting of the I-th band-pass filter characteristic 120i in this embodiment. The audio signal S that changes over time within a certain range around the center frequency fi is expressed as S=F(t,fi) by the time change t and the function F(). The maximum value Smax(fi) that the function F() can take at the center frequency fi is expressed as Smax(fi)=max(F(t,fi)). In addition, the minimum value Smin(fi) that the function F() can take at the center frequency fi is expressed as Smin(fi)=min(F(t,fi)). The width di of the frequency change that can be taken at the center frequency fi is calculated from the difference between the maximum value Smax(fi) and the minimum value Smin(fi) as di=Smax(fi)-Smin(fi). The width Di is made to match the frequency change width di, so it is expressed as Di=di. The interval Li(i+1) is set to 1 / 2 of the frequency change width di, so Li(i+1)=Di / 2. Also, the peak value Pi is expressed as Pi=C / Di, where C is a constant.

[0020] In the multiple bandpass filters 20 shown in FIG. 1, the bandpass filter characteristics 120 are determined in order from the bandpass filter 20 with the lowest frequency. For example, D1, L12, and P1 are determined based on f1 and d1 for the first bandpass filter 20a. f2 is determined by adding L12 to f1. D2, L23, and P2 are determined based on f2 and d2 for the second bandpass filter 20b. f3 is determined by adding L23 to f2. D3, L34, and P3 are determined based on f3 and d3 for the third bandpass filter 20c. The bandpass filter characteristics 120 for the subsequent bandpass filters 20 are determined in the same manner. Here, values ​​assumed in advance are used for the change widths d1, d2, d3, . . . In this way, each bandpass filter characteristic 120 is determined by the change width di of the frequency of the audio signal. Also, even if the change width di of the frequency differs depending on the center frequency fi, each change width di can be set independently. In the above description, the bandpass filter characteristic 120 is described as an example in which the width Di and the frequency change width di are the same, but the present invention is not limited to this. The bandpass filter characteristic 120 may be determined such that the frequency change width di is an integer multiple of the width Di. For example, the frequency change width di may be determined to be twice the width Di.

[0021] Figures 3(a)-(b) show the feature quantities output from each bandpass filter 20. Figure 3(a) shows a signal 130 input to a plurality of bandpass filters 20 having bandpass filter characteristics 120 shown in the same manner as in Figure 2. The horizontal axis indicates frequency, and the vertical axis indicates level. The signal 130 has a constant level even if the frequency changes.

[0022] FIG. 3(b) shows the feature quantities 140 output from each of the band pass filters 20 when a signal 130 is input to each of the band pass filters 20. The I-th feature quantity 140i is a value output from the I-th band pass filter 20i, and the I+1-th feature quantity 140i+1 is a value output from the I+1-th band pass filter 20i+1. The I+2-th feature quantity 140i+2 is a value output from the I+2-th band pass filter 20i+2, and the I+3-th feature quantity 140i+3 is a value output from the I+3-th band pass filter 20i+3. When a signal 130 of a constant level is input to the band pass filters 20 in this way, all the feature quantities 140 output from each band pass filter 20 have the same value. In this way, the feature quantity 140 is obtained by reducing the number of dimensions of a signal having frequency information such as a voice signal.

[0023] Up to this point, the bandpass filter characteristic 120 has been described in general terms, but here, the bandpass filter characteristic 120 will be described in detail. FIG. 4 shows the configuration of a wind power generator 200 to be determined. The wind power generator 200 includes a support pole 202, a first blade 210a, a second blade 210b, and a third blade 210c, which are collectively referred to as blades 210, and a rotation center 212. Here, the determination target is not limited to a wind power generator, but may be any device having rotating blades. The support pole 202 is a pole that extends in the vertical direction, and supports a plurality of blades 210 at its upper portion. The plurality of blades 210 can rotate around the rotation center 212. Here, it is assumed that a scratch 220 exists on the first blade 210a. In addition, a microphone 110 is attached to the support pole 202. The microphone 110 captures the sound generated by the blades 210 of the wind power generator 200, and outputs an audio signal generated based on the captured sound to an input section 10 (not shown) of the feature extraction and determination device 100 (not shown).

[0024] Here, the distance from microphone 110 to center of rotation 212 is denoted as "h", the length of blade 210 is denoted as "L", and the length from center of rotation 212 to flaw 220 is denoted as "l". Furthermore, the unit time of blade 210, for example the number of rotations per minute, is denoted as "R". Furthermore, the speed of sound is denoted as "v", and the speed between flaw 220 and microphone 110 is denoted as "vs".

[0025] In the wind power generator 200, a whistling sound is generated by a flaw 220 in the first blade 210a. The time change of the audio signal within a certain range of the center frequency generated in the whistling sound is due to the Doppler effect caused by the rotation of the blade 210. The speed vs between the flaw 220 and the microphone 110 is calculated by partially differentiating the distance between the flaw 220 and the microphone 110 with respect to time as follows. Here, it is noted that the maximum value of the length l from the rotation center 212 to the flaw 220 is the length L of the blade 210, and the maximum value of the number of rotations per minute R is predetermined, and the speed vs between the flaw 220 and the microphone 110 when the frequency change width due to the Doppler effect is the largest is calculated. The case where the frequency change width due to the Doppler effect is the largest is when the flaw 220 is located at the outer end of the blade L. By calculating the speed vs in this way, it is possible to suitably extract the feature amount even when the flaw 220 is closer to the rotation center 212 than the outer end of the blade L and the frequency change width due to the Doppler effect is smaller.

number

[0026] When a whistling sound is generated at the flaw 220 at the outer end of the blade L, the frequency f of the whistling sound subjected to the Doppler effect is calculated as follows. f = F(t,fi) = fi v / (v + vs)

[0027] As described above, if the maximum value Smax(fi) of the audio signal input to the I-th band-pass filter 20i is represented as max(F(t,fi)) and the minimum value Smin(fi) of the audio signal input to the I-th band-pass filter 20i is represented as min(F(t,fi)), the frequency change width di is calculated as follows: di=Smax(fi)-Smin(fi)=max(F(t,fi))-min(F(t,fi)).

[0028] That is, when the audio signal is a sound generated by the blade 210 of the wind power generator 200, the frequency change width di is determined based on the number of rotations R of the blade 210, the length L of the blade 210, and the distance h from the microphone 110 to the center of rotation 212. Using the frequency change width di thus determined, the bandpass filter characteristic 120 is calculated as described above. The bandpass filter characteristic 120 thus calculated is set for each bandpass filter 20.

[0029] 5(a)-(c) show an overview of the processing for the wind power generator 200. FIG. 5(a) shows the spectrum of an audio signal received by the input unit 10. It is known that time domain data such as an audio signal can be converted into frequency domain information by Fourier transforming it. Therefore, the spectrum of the audio signal can be derived, for example, by using a fast Fourier transform (FFT). The horizontal axis indicates frequency, and the vertical axis indicates level. The spectrum of the audio signal is 3×10 3 [Hz]: 300 non-whistle noises with increasing levels at frequencies around 10 4 The noise 300 other than the whistling sound is not the whistling sound 310 but an unexpected noise. The whistling sound 310 is caused by the Doppler effect caused by the flaw 220 in the first blade 210a.

[0030] FIG. 5(b) shows the feature quantity output from a known Mel filter bank by inputting the speech signal in FIG. 5(a) to the Mel filter bank. The horizontal axis shows the frequency, and the vertical axis shows the output value. The Mel filter bank is a 10 2 ~2×10 4 The feature values ​​are set to be output for 25 points in a frequency range of [Hz]. The feature values ​​output from the Mel filter bank include non-whistle noise 302 corresponding to non-whistle noise 300, and also include whistle sound 312 corresponding to whistle sound 310.

[0031] Fig. 5(c) shows the feature quantity output from the filter bank 12 by inputting the audio signal of Fig. 5(a) to the filter bank 12. The horizontal axis indicates frequency, and the vertical axis indicates output value. The feature quantity output from the filter bank 12 includes a whistling sound 322 corresponding to the whistling sound 310. However, the feature quantity output from the filter bank 12 does not include a component corresponding to noise 300 other than the whistling sound. In other words, by using the filter bank 12, it is possible to extract the whistling sound 310 without extracting the noise 300 other than the whistling sound. Return to Fig. 1.

[0032] The judgment unit 14 receives the features of the audio signal extracted by the filter bank 12. The judgment unit 14 judges whether the audio signal is normal or above normal based on a graph of the features. To explain the processing in the judgment unit 14, Fig. 6 is also used here. Fig. 6 shows an overview of the processing by the judgment unit 14. "Index" indicates the number arranged in order of center frequency of the bandpass filter, and the smaller the "index" is, the lower the frequency of the bandpass filter, and the larger the "index" is, the higher the frequency of the bandpass filter. "Value" indicates the numerical data of Fig. 5(c).

[0033] The determination unit 14 (i) calculates the difference from the data of the immediately preceding index. The determination unit 14 (ii) calculates the average of the differences, and sets this as the slope of the overall average of the differences. In FIG. 6, the average of the differences is approximately -1.916, and the overall average slope is "negative." The determination unit 14 (iii) uses the overall average slope of the differences as a reference, and assigns "E" to differences whose sign is different from the reference slope. In FIG. 6, since the reference slope is "negative," "E" is assigned to differences whose sign is "positive." This corresponds to extracting a portion whose value increases as the index increases.

[0034] The determination unit 14 (iv) extracts the value of the index immediately preceding the index to which "E" is added. Here, if the indexes to which "E" is added are consecutive, the value of the index immediately preceding only the smallest index is extracted. In FIG. 6, the indexes to which "E" is added are "3" to "5", "7", and "22". Therefore, in this case, for indexes "3", "7", and "22", the values ​​of the index immediately preceding them are extracted. This corresponds to obtaining the value of the portion where the value begins to increase as the index increases.

[0035] The determination unit 14 (v) gives a circle to the portion where the consecutive indexes with "E" are 1 to 2. In FIG. 6, "E" is given consecutively to indexes "3" to "5", but the consecutive indexes are 3, so no circle is given. In contrast, indexes "7" and "22" are given a circle because the indexes with "E" are not consecutive, that is, the consecutive indexes are 1. This corresponds to the design of the bandpass filter characteristic 120, in which the width Di is made to match the frequency change width di, that is, when a whistling sound is present, the feature quantity to be extracted does not occur two or more consecutively, so that other features are excluded. Here, the condition of (v) in the process of the determination unit 14 can be changed based on the relationship between the width Di and the frequency change width di. For example, when the frequency change width di is twice the width Di, the determination unit 14 gives a circle to the portion where the consecutive indexes with "E" are 3 to 4.

[0036] The determination unit 14 extracts indexes with a circle (vi) that have a value extracted in (iv) that is smaller than the value of (iv) of the index smaller than the index. For example, in FIG. 6, the indexes with a circle (v) are "7" and "22", and the indexes with values ​​extracted in (iv) are "2", "7", and "22". In this case, the value of (iv) of the circled index "7" is "9", and the value of (iv) of the index "3" is "2", which means that the value of (iv) of the index "7" is larger than the value of (iv) of the index "3", so the condition of (vi) is not satisfied. The value of (iv) of the circled index "22" is "-26", and the values ​​of (iv) of the indexes "3" and "7" are "2" and "9", which means that the value of (iv) of the index "22" is smaller than the value of (iv) of the indexes "3" and "7", so the condition of (vi) is satisfied. As a result, the determination unit 14 determines that the whistling sound 310 exists at the frequency of index "22." This corresponds to determining whether the change extracted by (v) is part of a larger change or a small change occurring only around that index.

[0037] When the reference gradient calculated in (ii) is "positive", the judgment unit 14 performs processing by appropriately switching the signs in the conditions (iii) to (vi). In this way, the whistling sound 310 can be extracted whether the reference gradient is "positive" or "negative".

[0038] 6, there is only one whistling sound 310, but this is not limiting. Depending on the state of the scratches 220 on the blade 210, there may be multiple whistling sounds at different frequencies. In any case, the determination unit 14 can extract the whistling sound 310 by performing processing under conditions (i) to (vi).

[0039] This configuration can be realized in hardware terms by the CPU, memory, and other LSIs of any computer, and in software terms by programs loaded into memory, but here we depict functional blocks realized by the cooperation of these. Therefore, it will be understood by those skilled in the art that these functional blocks can be realized in various forms by hardware alone, software alone, or a combination of both.

[0040] According to this embodiment, the width Di, interval Li(i+1), and peak value Pi are set based on the width di of the frequency change that can occur at the center frequency fi, so that it is possible to accurately extract features contained in an audio signal whose frequency may change periodically within a certain range due to a physical phenomenon. In addition, the width di of the frequency change is determined based on the number of rotations R of the blade, the length L of the blade, and the distance h from the microphone to the center of rotation of the blade, so that it is possible to accurately extract a whistling sound from an audio signal generated by the blades of a wind power generator.

[0041] Moreover, since the bandpass filter characteristics are set based on the width di of the frequency change, it is possible to suitably extract features for detecting the whistling sound with high accuracy without detecting the Doppler shift. Moreover, since the length of the audio signal is equal to or longer than an integer multiple of the period, it is possible to suitably extract features for detecting the whistling sound with high accuracy. Moreover, since it is determined whether the audio signal is normal or not based on a graph of the features extracted in the filter bank, it is possible to detect the whistling sound with high accuracy.

[0042] The present invention has been described above based on the embodiments. These embodiments are merely illustrative, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component and each treatment process, and that such modifications are also within the scope of the present invention.

[0043] In this embodiment, the determination unit 14 detects a whistling sound. However, this is not limiting, and for example, the filter bank 12 does not detect a change called a Doppler shift, but allows a change to extract a feature amount, so that an abnormal sound other than a whistling sound may be detected. According to this modification, the range of application of this embodiment can be expanded.

[0044] The determination unit 14 in this embodiment detects the whistling sound based on a graph of the feature amount. However, this is not limiting, and the determination unit 14 may detect the whistling sound by executing machine learning using the feature amount extracted in the filter bank 12. This modification can improve the degree of freedom of the configuration of this embodiment. [Explanation of symbols]

[0045] 10 input unit, 12 filter bank, 14 decision unit, 20 band pass filter, 22 synthesis unit, 100 feature extraction / decision device, 110 microphone, 120 band pass filter characteristic, 130 signal, 140 feature, 200 wind turbine, 202 support pole, 210 blade, 212 center of rotation, 220 scratch.

Claims

1. Receiving an audio signal that may include components whose frequency changes periodically within a certain range; inputting the received speech signal into a filter bank which is an array of a plurality of band pass filters, and extracting a feature of the speech signal; a width Di of the I band-pass filter, an interval Li(i+1) between the center frequency fi of the I band-pass filter and the center frequency fi+1 of the I+1 band-pass filter, and a peak value Pi of the I band-pass filter are set based on a width di of the frequency change that can occur at a center frequency fi of the I band-pass filter.

2. 2. The feature extraction method according to claim 1, wherein, when the audio signal is a sound generated by a rotating blade, the range di of the frequency change is determined based on the number of rotations R of the blade, the length L of the blade, and the distance h from a microphone that acquires the audio signal to the center of rotation of the blade.

3. The velocity between the blade cut and the microphone, vs, is calculated as follows: [0010] If the voice signal input to the I-th bandpass filter is denoted as S(fi) = F(t,fi) and the sound speed is denoted as v, the frequency f affected by the Doppler effect is calculated as follows: f=F(t,fi)=fi・v / (v+vs) If the maximum value Smax(fi) of the audio signal input to the Ith band pass filter is represented as max(F(t,fi)) and the minimum value Smin(fi) of the audio signal input to the Ith band pass filter is represented as min(F(t,fi)), the frequency change width d is calculated as follows: di=Smax(fi)-Smin(fi)=max(F(t,fi))-min(F(t,fi)) The feature extraction method according to claim 2 .

4. 4. A method for determining whether the audio signal is normal or above normal based on a graph of the feature extracted by the feature extraction method according to claim 1.

5. An input unit that receives an audio signal that may include components whose frequency changes periodically within a certain range; a filter bank, which is an array of a plurality of band pass filters, to which the received voice signal is input and which extracts a feature of the voice signal; In the filter bank, when one bandpass filter of the plurality of bandpass filters is designated as an I-th bandpass filter and a bandpass filter adjacent to the I-th bandpass filter in the frequency domain is designated as the I+1-th bandpass filter, a width Di of the I-th bandpass filter, an interval Li(i+1) between the center frequency fi of the I-th bandpass filter and the center frequency fi+1 of the I+1st bandpass filter, and a peak value Pi of the I-th bandpass filter are set based on a width di of the frequency change that can occur at a center frequency fi of the I-th bandpass filter.

Images