JP4760179B2 - Voice feature amount calculation apparatus and program - Google Patents

Description

  The present invention relates to a signal generation technique for use in a speech recognition apparatus.

  Various speech recognition apparatuses that automatically perform speech recognition have been proposed. In general, the speech recognition apparatus is based on the similarity between the speech feature amount corresponding to various words stored in advance and the speech feature amount uttered by the speaker. Recognize

  Various methods for calculating the feature amount of speech used in speech recognition devices have been proposed. As one of those calculation methods, the filter bank output value obtained by filtering the speech spectrum with a filter provided for each of a plurality of frequency bands is transformed using discrete cosine transform or discrete inverse Fourier transform, and the speech There is a method for calculating a coefficient sequence indicating a feature amount. MFCC (Mel-Frequency Cepstrum Coefficient, Mel Frequency Cepstrum Coefficient) is an example of a coefficient sequence calculated by such a method and is widely used. The mechanism of speech recognition according to the prior art will be described below by taking as an example the case of using MFCC obtained by discrete cosine transform.

  FIG. 7 is a block diagram showing the configuration of a speech recognition system 9 according to the prior art. The voice recognition system 9 includes a voice signal generation device 90 that converts a voice of a speaker into a voice signal, a voice feature value calculation device 91 that calculates a MFCC using a voice signal generated by the voice signal generation device 90, a voice feature value A speech recognition device 92 that performs speech recognition using the MFCC calculated by the calculation device 91 is provided.

  The audio signal generation device 90 includes an audio signal generation unit 901 that collects audio and converts it into an audio signal, and an audio signal generated by the audio signal generation unit 901 includes, for example, an interval having an amplitude value equal to or greater than a predetermined threshold. As shown in FIG. The speech signal of the speech segment extracted by the speech segment extraction unit 902 is divided into, for example, 40 ms long frames, and then output from the speech signal generation device 90 to the speech feature value calculation device 91.

  The audio feature quantity calculation device 91 reduces the high-frequency noise due to frame division by multiplying the audio signal in units of frames received from the audio signal generation device 90 while sliding a time window function such as a Hamming window in the time axis direction. A windowing processing unit 911 that generates a voice signal, and an FFT (Fast Fourier Transform) process is performed on the audio signal of each frame subjected to the windowing process by the windowing processing unit 911 to calculate the spectrum of the voice signal An FFT processing unit 912 and a filter group called a mel scale band filter (described later) filter the spectrum calculated by the FFT processing unit 912 to calculate an index value indicating the power of the frequency component for each of the plurality of frequency bands. Mel scale band filter processing unit 9 13. A logarithmic value calculation unit 914 that calculates the logarithmic value of each index value calculated by the mel scale band filter processing unit 913, and a set of logarithmic values calculated by the logarithmic value calculation unit 914 are subjected to discrete cosine transform (described later). Accordingly, a discrete cosine transform processing unit 915 for calculating the MFCC is provided.

The mel-scale bandpass filter has a multiplier at the center frequency of 1 adjacent to each of a plurality of center frequencies arranged at equal intervals on the mel frequency axis obtained by converting the linear frequency axis by the following equation (1). It is a collection of filters whose multipliers are linear and change so that the multiplier at the center frequency of the filter to be 0 becomes zero.

FIG. 8 is a graph showing a melscale bandpass filter. As shown in FIG. 8, for example, the center frequency f _k filter 95 (Hz) takes the multiplier 1 at the center frequency f _k (Hz), the low-frequency side of the adjacent filter center frequency f _k-1 (Hz) And a triangular shape having a multiplier of 0 at the center frequency f _{k + 1} (Hz) of adjacent filters on the high frequency side. In this case, each triangular shape in FIG. 8 is called a filter bank.

  By the way, the purpose of converting the linear frequency axis to the Mel frequency axis is to consider human auditory characteristics that are less sensitive to pitch changes in the high frequency band than in the low frequency band. This is because it is possible to indicate the distance between frequencies along the auditory sense.

The melscale band filter processing unit 913 multiplies the spectrum calculated by the FFT processing unit 912 by each filter bank of the mel scale band filter and adds the spectrum power to the spectrum included in the frequency band covered by each filter bank. As an index value, a filter bank output value r _k (where k is a filter bank number) is calculated. Hereinafter, the number of filter banks is L.

The logarithmic value calculation unit 914 calculates each logarithmic value R _k of the filter bank output value r _k calculated by the mel scale band filter processing unit 913. The discrete cosine transform processing unit 915 calculates a MFCC that is a coefficient sequence by performing a discrete cosine transform on the logarithmic value R _k of the filter bank output value calculated by the logarithmic value calculation unit 914 according to the following equation (2). However, C _i in equation (2) shows the i-th order coefficient in MFCC.

Here, an effective numerical value is obtained with i as an upper limit of about 1/2 of L. For example, if the number of filter banks is 12, C ₁ , C ₂ ,..., C ₆ can be obtained as effective MFCCs. Audio feature amount calculation unit 91 outputs the C _i group that has been calculated as described above to the speech recognition device 92.

The voice recognition device 92 has two operation modes, a learning mode and a recognition mode. In the learning mode, the speech recognition device 92 receives the C _i group for each frame of the speech signal indicating the speech utterance's from the audio feature amount calculation unit 91, a C _i group for each frame to a speech signal indicating a series of sound , And sequentially stored in the database 921 in association with words pronounced by the speaker. Accordingly, the coefficient sequence group showing a feature amount corresponding to the word, for example "good morning" in database 921, C _i group is to be stored in chronological order in accordance with the number of frames the audio signal of "good morning" . Below, a series gatherings when the C _i group referred to as the "C _i group columns". The speaker speaks various words sequentially, and inputs the pronounced words into the speech recognition device 92 by operating means such as a keyboard (not shown) connected to the speech recognition device 92, for example. C _i group row corresponding to the word can be sequentially stored.

On the other hand, the voice recognition device 92 includes a DP matching unit 922 and a determination unit 923 that perform processing in the recognition mode. DP matching unit 922 calculates a distance indicating a similarity between each of the C _i groups string stored in C _i group columns and database 921 receives from the audio feature amount calculation unit 91 by the DP (Dynamic Programming) matching method . The determination unit 923 determines whether the distance calculated relates to any of the C _i groups string stored in the database 921 is the shortest. Furthermore, the determination unit 923, the distance the words are stored in the database 921 in association with the C _i group string is determined to be the shortest, and specified as the word was uttered by the speaker, the identified word Data to be shown is transmitted to another device or notified as a message to the user.
The above is the mechanism in which speech recognition is performed by the speech recognition system 9 according to the conventional technology.

  By the way, in the speech recognition system 9, speech recognition is performed with the expected accuracy if the sound space where the speaker is placed is low noise regardless of whether the speech recognition device 92 is in the learning mode or the recognition mode. However, in general, there is an environmental noise that cannot be ignored in the sound space where the speaker speaks. Therefore, the MFCC generated by the speech recognition system 9 indicates a feature amount of sound in which environmental noise is mixed with the voice of the speaker. As a result, in the speech recognition system 9, speech recognition is not always performed with the expected accuracy.

In order to solve the above problem, it is conceivable to perform a noise reduction process called a spectral subtraction on the audio signal, for example. Spectral subtraction is a technique for extracting the spectrum of a sound signal indicating sound by subtracting the spectrum of a sound signal indicating environmental noise from the spectrum of a sound signal indicating sound mixed with sound and environmental noise. For example, Patent Document 1 discloses a technique for detecting a speech section from a sound signal using spectrum subtraction.
JP 2000-47696 A

  Many noise reduction processes such as spectral subtraction have excellent effects and require a large amount of calculation, and may be difficult to implement in a device with severe resource constraints such as a portable terminal device.

  In view of the above situation, the present invention makes it possible to reduce the influence of environmental noise included in a coefficient sequence by a simple and low-load process in a system that calculates a coefficient sequence indicating a feature amount of speech. It aims to provide a means.

In order to achieve the above object, the present invention provides a spectrum calculation unit that calculates a spectrum of the audio signal from an audio signal, and a spectrum calculated by the spectrum calculation unit according to each of a plurality of predetermined frequency bands. Filter means for calculating an index value indicating the power of the frequency component in the frequency band included in the audio signal for each of the plurality of frequency bands by performing filtering, and each of the plurality of frequency bands A logarithmic value calculating means for calculating a logarithmic value of each of the plurality of index values calculated by the filter means; and a minimum value of the plurality of logarithmic values calculated by the logarithmic value calculating means for each of the plurality of frequency bands. m, where M is the maximum value, constant p , constant q (where m ≦ p <q ≦ M) and constant n (where n> 1) Y ≧ m when the output value y for the value x is (a) x = m, (b) y ≦ M when x = M, and (c) y for x in the range of m ≦ x ≦ M. In accordance with the following expression (3) that satisfies the condition that the rate of change of y with respect to x is always positive in the range of (d) p ≦ x ≦ q. There is provided a speech feature quantity calculation device comprising: a conversion means for calculating a numerical value sequence indicating a feature quantity of speech indicated by the speech signal by converting each of the index values.

  According to such an audio feature amount calculation device, a logarithm value of a small value related to a filter bank including a relatively large amount of environmental noise feature amount components is converted to be smaller than that of speech, and the environmental noise feature amount component is not much included. The logarithm of a large value for a filter bank that is not converted is converted to a larger value, and the converted logarithm does not exceed the range of the minimum and maximum values of the logarithm before conversion, and is included in the logarithm of the small value Since the component of the voice feature value is not underestimated, for example, a numerical sequence that provides a desired recognition result when used for voice recognition is calculated.

Further, the present invention provides a spectrum calculation unit that calculates a spectrum of the audio signal from an audio signal, and a filter process corresponding to each of a plurality of predetermined frequency bands on the spectrum calculated by the spectrum calculation unit. Filter means for calculating an index value indicating the power of the frequency component in the frequency band included in the audio signal for each of the plurality of frequency bands; and the filter means for each of the plurality of frequency bands. Logarithmic value calculation means for calculating the logarithmic value of each of the plurality of index values; m for the minimum value of the plurality of logarithmic values calculated by the logarithmic value calculation means for each of the plurality of frequency bands; , Constant p, constant q (where m ≦ p <q ≦ M), constant a (where a> 0) and constant c (where c> 0) When the output value y with respect to the input value x is (a) x = m, y ≧ m, (b) when x = M, y ≦ M, and (c) in the range of m ≦ x ≦ M, In accordance with the following equation (2) that satisfies the condition that the rate of change of y is always positive in the range of (d) p ≦ x ≦ q, the rate of change of y is always 0 or more. There is provided a speech feature quantity calculation device comprising: a conversion means for computing a numerical value sequence indicating a feature quantity of speech indicated by the speech signal by converting each of a plurality of index values.

  A preferred example of the filter used by the filter means is a mel scale band filter. In this case, the speech feature quantity calculation device may be configured to include coefficient sequence calculation means for calculating a mel frequency cepstrum coefficient sequence by performing discrete cosine transform on the numerical value sequence calculated by the conversion means.

  In addition, the present invention provides a program that causes a computer to execute processing performed by the above-described audio feature amount calculation apparatus.

[Embodiment]
FIG. 1 is a block diagram showing a configuration of a speech recognition system 1 according to an embodiment of the present invention. The speech recognition system 1 is common in many respects to the speech recognition system 9 according to the above-described prior art, and only different points will be described below. In FIG. 1, components common to the speech recognition system 1 and the speech recognition system 9 are denoted by the same reference numerals as those in FIG.

  The voice recognition system 1 includes a voice feature quantity calculation device 11 instead of the voice feature quantity calculation device 91 of the voice recognition system 9. The voice recognition system 1 also includes a keyboard 12 that is used by the user to give instructions to the voice feature quantity calculation device 11. The keyboard 12 includes a plurality of keys, and outputs a signal corresponding to the key pressed by the user to the sound feature quantity calculation device 11. For example, a mouse pointer may be used instead of the keyboard 12.

The speech feature quantity calculation device 11 includes a conversion unit 101 interposed between a logarithmic value calculation unit 914 and a discrete cosine transform processing unit 915 in addition to the components included in the speech feature quantity calculation device 91. Conversion unit 101, the filter bank output value is calculated by the logarithm calculator 914 r _k (Here, k filter bank number) receives logarithmic value R _k of the above-described logarithmic values R _k received as an input value x By substituting into the equation (3), the output value y corresponding to the logarithmic value R _k is calculated. Hereinafter referred to as modified logarithmic value gamma _k output value y corresponding to the logarithm R _k.

  In addition, the audio feature quantity calculation device 11 includes a designation unit 102 that designates parameters for the conversion unit 101 in accordance with a signal output from the keyboard 12 in accordance with a user operation. In this case, the parameter designated by the designation unit 102 is the constant n in the expression (3).

The discrete cosine transform processing unit 915 in the speech feature quantity calculation device 11 is a logarithmic value group R ₁ , R ₂ ,..., R _L (where L is the total number of filter banks) generated by the logarithmic value calculation unit 914. Instead, by receiving the modified logarithmic value groups γ ₁ , γ ₂ ,..., Γ _L calculated by the conversion unit 101, the received modified logarithmic value group is subjected to discrete cosine transform according to the above-described equation (2), C _i group, that is, MFCC is calculated. However, since the MFCC in the speech recognition system 1 is generated using a logarithmic value group having characteristics different from those of the MFCC in the prior art, the MFCC has characteristics different from those of the MFCC in the prior art.

FIG. 2 is a graph in which the function represented by Expression (3) is drawn at coordinates with the horizontal axis representing the input value x and the vertical axis representing the output value y. However, in FIG. 2, graph 15, graph 16 and graph 17 show graphs in the case of n = 1.5, n = 3.0 and n = 4.5, respectively. The modified logarithmic value γ _k for the logarithmic value R _k in the case is shown.

As shown in FIG. 2, the conversion unit 101 substitutes the logarithmic value R _k as the input value x into the equation (3) to calculate the modified logarithmic value γ _k as the output value y. The modified logarithmic value γ _k has the following characteristics.
(A) The magnitude relationship between the input values is always maintained in the magnitude relationship between the output values.
(B) In the output value in the region where the input value is large (region near the left side of x = M), the size of the input value is substantially maintained.
(C) In the region where the input value is small (region near the right side of x = m) or the region where the input value is medium, the output value for the input value in the region where the input value is large (region near the left side of x = M) There is a wide range in which the reduction range is larger than the reduction range.
(D) The output value is always within the range of the minimum value and the maximum value of the input value.

Component of the feature quantities of the environmental noise, the logarithmic value group R _1, R _2, · · ·, of R _L, contains many by what that value is small. This is because the power of the ambient noise spectrum is generally smaller than the power of the voice spectrum for the entire frequency band. Therefore, in the modified logarithmic value groups γ ₁ , γ ₂ ,..., Γ _L having the characteristics as described in (b) and (c) above, the logarithmic value groups R ₁ , R ₂ ,. Compared to the case of _L, the component of the feature amount of the environmental noise is evaluated to be small, and the component of the speech feature amount is not evaluated to be very small. As a result, variations logarithm group _{γ 1, γ 2, ···,} C i group to be calculated using the gamma _L, i.e. MFCC also logarithm groups R _1, R _2, · · ·, a R _L using C _i group that is calculated, that is, compared to MFCC, the indicators including fewer components of the feature amount of environmental noise.

By the way, in order to generate the modified logarithmic value groups γ ₁ , γ ₂ ,..., Γ _L having the characteristics as described in (b) and (c) above, for example, conversion according to the following equation (5) Can also be considered.

However, according to Equation (5), logarithmic value group R _1, R _2, · · ·, logarithm of the minimum value of R _L is converted to 0, although logarithm value not zero close to the minimum value , Converted to a fairly small value. As a result, not only the environmental noise feature quantity but also the voice feature quantity related to the frequency band in which the spectrum power is small is underestimated. As a result, a desired speech recognition result may not be obtained depending on the MFCC calculated using the modified logarithmic value group obtained by the conversion according to the equation (5).

On the other hand, the modified logarithmic value groups γ ₁ , γ ₂ ,..., Γ _L calculated by the conversion unit 101 have the above-mentioned feature (d), and thus do not cause the above-described adverse effects.

  In the speech recognition system 1, the parameter n can be changed by the user giving an instruction to the speech feature quantity calculation device 11 using the keyboard 12. As a result, the user can easily select a function that seems to be desirable from functions having different characteristics as illustrated in FIG. 2, and can cause the audio feature quantity calculation device 11 to generate different specific MFCCs. it can. Therefore, it is possible to generate a more suitable MFCC according to the environmental noise situation. Note that the parameters used in the learning mode are stored in, for example, the conversion unit 101, and the same parameters as those used in the learning mode are used in the recognition mode.

As described above, according to the speech recognition system 1 according to the embodiment of the present invention, the MFCC having a preferable characteristic that the feature amount related to the environmental noise is not included but the feature amount related to the speech is not underestimated is calculated. Is done. As a result, a more accurate speech recognition result can be obtained as compared with the case of the speech recognition system 9 according to the prior art. At this time, the speech feature quantity calculation device 11 calculates the modified logarithmic value γ _k by substituting the logarithmic value R _k for the function shown in the equation (3), as compared with the speech feature quantity calculation device 91 according to the prior art. Only processing has been added. Therefore, the speech feature quantity calculation device 11 can be realized even by a device with limited resources.

[Modification]
By the way, the conversion unit 101 in the embodiment described above has been described as converting the logarithmic value R _k to a modified logarithmic value gamma _k by equation (3) is not limited thereto, it satisfies the condition different functions: converting unit It can be used in 101 conversion.
When the minimum value of the logarithmic value groups R ₁ , R ₂ ,..., _L is m and the maximum value is M, the input value x is related to the constant p and the constant q (where m ≦ p <q ≦ M). When the output value y for (a) is x = m, y ≧ m.
(B) When x = M, y ≦ M.
(C) In the range of m ≦ x ≦ M, the rate of change of y with respect to x is always 0 or more.
(D) In the range of p ≦ x ≦ q, the rate of change of y with respect to x is always positive.

The modified logarithmic value groups γ ₁ , γ ₂ ,..., Γ _L calculated in the conversion by the function satisfying the above conditions have the above-described features (a) to (d). An example of a function that satisfies the above condition is, for example, the above expression (4). Equation (4) is obtained by deforming a logistic curve using the minimum value m and the maximum value M. 3 and 4 are graphs in which the function represented by Expression (4) is drawn at coordinates with the horizontal axis representing the input value x and the vertical axis representing the output value y. However, in FIG. 3, graph 21, graph 22 and graph 23 show the change in shape when constant a is fixed at a = 10 and constant c is changed to c = 20, c = 100 and c = 400, respectively. In FIG. 4, the graph 24, the graph 25, and the graph 26 have the shapes when the constant c is fixed at c = 100 and the constant a is changed to a = 20, a = 10, and a = 7, respectively. It shows a change. As described above, when the conversion unit 101 performs the conversion according to the equation (4), the user designates the parameter a and the parameter c with respect to the voice feature amount calculation device 11 using the keyboard 12, and thus a more desirable conversion result. Can be selected.

  Furthermore, the conversion unit 101 stores a conversion table as shown in FIG. 5 in advance, instead of performing conversion from input values to output values using functions such as those shown in equations (3) and (4). Alternatively, the same conversion may be performed according to the conversion table. The input value x and the input value y included in the conversion table used by the conversion unit 101 are a set of numerical values that satisfy the conditions (a) to (d). In addition, a set of numerical values included in the conversion table is created on the assumption that, for example, the minimum value m = 0 of the input value and the maximum value M = 1 of the input value. Hereinafter, the conversion table created with m = 0 and M = 1 is referred to as a “reference conversion table”. FIG. 6 is a graph in which the input value x and the input value y of the conversion table shown in FIG. 5 are plotted.

As described above, since the reference conversion table is for m = 0 and M = 1, the conversion unit 101 does not use the reference conversion table as it is, but the received logarithmic value groups R ₁ , R ₂ ,. .., The reference conversion table is converted and used according to the minimum values m and M of _RL . Specifically, the conversion unit 101 creates a value obtained by multiplying the input values x and y of the reference conversion variable by (M−m), respectively, and adding m to the input values x and y, respectively. Using the conversion table, the logarithmic value group R _k is converted into a modified logarithmic value γ ₁ . Since the input value x and the output value y included in the conversion table are discrete values, the conversion unit 101 interpolates numerical values by linear interpolation or the like as necessary when calculating the output value y for the input value x.

  In the above-described embodiment, the FFT processing is performed when calculating the spectrum of the audio signal, but other methods such as discrete Fourier transform processing may be used instead. In the above-described embodiment, the mel scale is used, but another frequency axis such as a bark scale may be used instead. Furthermore, instead of the mel scale band filter, another type of filter group that performs filter bank output may be used.

  In the above-described embodiment, the MFCC is calculated using the discrete cosine transform. However, instead of the discrete cosine transform, the speech may be converted using another type of orthogonal transform such as a discrete inverse Fourier transform. A coefficient sequence indicating the feature amount may be calculated.

  The audio feature quantity calculation device 11 may be realized by dedicated hardware, or may be realized by causing a general-purpose computer capable of inputting / outputting sound signals to execute processing according to an application program. When the speech feature quantity calculation device 11 is realized by a general-purpose computer, each component of the speech feature quantity calculation device 11 includes a CPU (Central Processing Unit) included in the general-purpose computer and a DSP (Digital Signal Processor) that operates under the control of the CPU. ) Is realized as a function of a general-purpose computer by performing processing according to each module included in the application program in parallel.

It is a block diagram which shows the structure of the speech recognition system concerning embodiment of this invention. It is a graph of the function which the conversion part concerning embodiment of this invention uses. It is a graph of the function which the conversion part concerning embodiment of this invention uses. It is a graph of the function which the conversion part concerning embodiment of this invention uses. It is a conversion table which the conversion part concerning embodiment of this invention uses. It is a graph of the conversion table which the conversion part concerning embodiment of this invention uses. It is a block diagram which shows the structure of the speech recognition system concerning a prior art. It is a graph which shows a mel scale bandpass filter.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 * 9 ... Voice recognition system, 11.91 ... Voice feature-value calculation apparatus, 12 ... Keyboard, 90 ... Voice signal generation apparatus, 92 ... Voice recognition apparatus, 101 ... Conversion part, 102 ... Designation part, 901 ... Voice signal generation , 902 ... utterance section extraction unit, 911 ... windowing processing unit, 912 ... FFT processing unit, 913 ... melscale band filter processing unit, 914 ... logarithmic value calculation unit, 915 ... discrete cosine transform processing unit, 921 ... database 922... DP matching unit, 923.

Claims (6)

Spectrum calculating means for calculating the spectrum of the audio signal from the audio signal;
A frequency within the frequency band included in the audio signal is included for each of the plurality of frequency bands by performing a filtering process corresponding to each of the plurality of predetermined frequency bands on the spectrum calculated by the spectrum calculating unit. Filter means for calculating an index value indicating the power of the component;
Logarithmic value calculation means for calculating the logarithmic value of each of the plurality of index values calculated by the filter means for each of the plurality of frequency bands;
When the minimum value of the plurality of logarithmic values calculated by the logarithmic value calculation means for each of the plurality of frequency bands is m and the maximum value is M, a constant p , a constant q (where m ≦ p <q ≦ M ) And constant n (where n> 1) , y ≧ m when the output value y with respect to the input value x is (a) x = m,
(B) y ≦ M when x = M,
(C) In the range of m ≦ x ≦ M, the rate of change of y with respect to x is always 0 or more,
(D) In the range of p ≦ x ≦ q, by converting each of the plurality of index values according to the following equation (1) that satisfies the condition that the rate of change of y with respect to x is always positive : And a conversion means for calculating a numerical string indicating the feature amount of the voice indicated by the voice signal.

Spectrum calculating means for calculating the spectrum of the audio signal from the audio signal;
A frequency within the frequency band included in the audio signal is included for each of the plurality of frequency bands by performing a filtering process corresponding to each of the plurality of predetermined frequency bands on the spectrum calculated by the spectrum calculating unit. Filter means for calculating an index value indicating the power of the component;
Logarithmic value calculation means for calculating the logarithmic value of each of the plurality of index values calculated by the filter means for each of the plurality of frequency bands;
When the minimum value of the plurality of logarithmic values calculated by the logarithmic value calculation means for each of the plurality of frequency bands is m and the maximum value is M, a constant p , a constant q (where m ≦ p <q ≦ M ) , With respect to the constant a (where a> 0) and the constant c (where c> 0) , y ≧ m when the output value y with respect to the input value x is (a) x = m,
(B) y ≦ M when x = M,
(C) In the range of m ≦ x ≦ M, the rate of change of y with respect to x is always 0 or more,
(D) In the range of p ≦ x ≦ q, by converting each of the plurality of index values according to the following equation (2) that satisfies the condition that the rate of change of y with respect to x is always positive : And a conversion means for calculating a numerical string indicating the feature amount of the voice indicated by the voice signal.

Before SL filter means, audio feature amount calculating apparatus according to claim 1 or 2, characterized in that for calculating the index value by mel scale band filter. By discrete cosine transform calculated numerical sequence by prior Symbol conversion unit, the audio feature amount calculating apparatus according to claim 3, characterized in that it comprises a coefficient string calculation means for calculating a Mel-frequency cepstral coefficient string. A process of calculating a spectrum of the audio signal from the audio signal,
An index value indicating the power of the frequency component in the frequency band included in the audio signal for each of the plurality of frequency bands by performing filtering processing on the spectrum according to each of the predetermined frequency bands. A process of calculating
Processing for calculating logarithmic values of each of the plurality of index values for each of the plurality of frequency bands;
When the minimum value of the plurality of logarithmic values for each of the plurality of frequency bands is m and the maximum value is M, a constant p, a constant q (where m ≦ p <q ≦ M) and a constant n (where n > 1) , y ≧ m when the output value y with respect to the input value x is (a) x = m,
(B) y ≦ M when x = M,
(C) In the range of m ≦ x ≦ M, the rate of change of y with respect to x is always 0 or more,
(D) In the range of p ≦ x ≦ q, by converting each of the plurality of index values according to the following equation (1) that satisfies the condition that the rate of change of y with respect to x is always positive : And a program for causing a computer to execute a process of calculating a numerical string indicating a feature amount of a voice indicated by the voice signal.

A process of calculating a spectrum of the audio signal from the audio signal,
An index value indicating the power of the frequency component in the frequency band included in the audio signal for each of the plurality of frequency bands by performing filtering processing on the spectrum according to each of the predetermined frequency bands. A process of calculating
Processing for calculating logarithmic values of each of the plurality of index values for each of the plurality of frequency bands;
When the minimum value of the plurality of logarithmic values for each of the plurality of frequency bands is m and the maximum value is M, a constant p, a constant q (where m ≦ p <q ≦ M) , a constant a (where a > 0) and constant c (where c> 0) , y ≧ m when the output value y with respect to the input value x is (a) x = m,
(B) y ≦ M when x = M,
(C) In the range of m ≦ x ≦ M, the rate of change of y with respect to x is always 0 or more,
(D) In the range of p ≦ x ≦ q, by converting each of the plurality of index values according to the following equation (2) that satisfies the condition that the rate of change of y with respect to x is always positive : And a program for causing a computer to execute a process of calculating a numerical string indicating a feature amount of a voice indicated by the voice signal.

Images