JP2655902B2 - Voice feature extraction device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech feature extraction device that extracts a feature amount independent of a speaker or language from input speech.

<Conventional Technology> Conventionally, input speech is recognized by a speech recognition device in the following manner. That is, the feature extraction unit performs frequency analysis on the input speech signal to extract in advance the feature amounts of some phonemes to be recognized. Then, the feature amounts of the plurality of phonemes are stored in the storage unit as standard patterns of the respective phonemes. Next, each word is expressed as a sequence of those phoneme standard patterns, and this phoneme standard pattern sequence is stored in the storage unit for each word in association with the phoneme sequence of the word, and stored as a word dictionary. On the other hand, when an unknown voice is input, the feature extracting unit extracts the feature amount of the input voice for each frame as described above. Then, the similarity between the extracted feature amount of each unknown speech frame and the phoneme standard pattern stored in the storage unit is checked, and the phoneme having the phoneme standard pattern with the highest similarity is determined as the phoneme of the frame. I do. Similarly,
The phonemes of each frame are sequentially determined, and the unknown speech is represented as a phoneme sequence. Then, the similarity between the phoneme sequence from the unknown voice and the phoneme standard pattern sequence of each word in the word dictionary stored in the storage unit is checked, and the word having the phoneme standard pattern sequence with the highest similarity is determined. It is determined as a word of the input voice.

In addition, a method has been proposed in which an articulatory model that directly expresses the vocal tract shape is set based on the structure of articulatory organs, and the articulation state is estimated from speech waves by a model matching method (Shirai, Yoshida: “Speech Estimation of Articulation Parameters from Waves ”Transactions of the Institute of Electronics, Information and Communication Engineers, '78 / 5 Vol.J61-A No.5). In this method, it is assumed that a non-linear function model for converting articulatory parameters representing the position of articulatory organs into acoustic parameters related to the speech spectrum is provided, and conversely, the above-described articulatory problem is solved by solving a nonlinear optimization problem from the acoustic parameters. The parameters are determined (that is, the articulation parameters are determined so as to minimize the fitting error of the functional model), and the articulation state is estimated.

<Problems to be Solved by the Invention> However, the above-described conventional input speech recognition method has the following problems. In other words, in the method based on the phoneme standard pattern, the feature amount of the phoneme extracted by the feature extraction unit is the speaker's physiological difference (for example, vocal tract length) Of the vowels in a word, there is a problem that the vowels in a word also differ due to the effect of articulatory coupling by the preceding and succeeding phonological environments. That is, if speech recognition is performed using such phoneme features, even if the speech utters the same phoneme symbol, it may be determined to be a different phoneme and rejected or erroneously recognized. High recognition performance cannot be obtained. Such a problem occurs because speech recognition is performed using a feature amount of a phoneme whose feature amount varies depending on a speaker or a phoneme environment.

Therefore, in order to solve such a problem, it is necessary to extract a characteristic amount that does not change even if the speaker changes or the phonological environment changes, that is, a characteristic amount unique to the utterance of the voice independent of the speaker or the language. .

Further, in the method of estimating articulatory parameters from acoustic parameters using an articulatory model, uniqueness of solution and stability of convergence pose problems. Therefore, Shirai et al. Incorporated the constraints of the variation range of articulatory parameters and the continuity between analysis frames into the evaluation function, considered the characteristics of the function model as separation criteria for vocal tract characteristics, and set appropriate initial values. The solution is sought by setting a limit. However, since the method of finding a solution as described above is complicated, it takes a long time to process, and it can be applied only to a stable place such as the center of a vowel, and can be applied only to a specific speaker.

Therefore, an object of the present invention is to extract, by simple processing, an articulation position (a narrow position formed in the vocal tract), which is a characteristic amount unique to the utterance of a voice independent of a speaker or language. An object of the present invention is to provide an audio feature extraction device.

<Means for Solving the Problems> In order to achieve the above object, the present invention relates to a voice feature extraction device for performing frequency analysis of an input voice and extracting a voice feature from an obtained frequency component. A vowel / consonant section determining unit for determining a section and a consonant section;
A table for converting at least two frequency components of a vowel into articulation positions, the conversion table including element values relating to a plurality of single vowels whose utterance contents are known, and an articulation position of a target vowel whose utterance contents are known The vowel / consonant interval determination unit determines a position on the conversion table relating to a target vowel whose utterance content determined to be a vowel interval is known by the vowel / consonant interval determination unit. , Based on the two frequency components of the single vowel and the position on the conversion table, and using the obtained element value at the position on the conversion table to calculate the articulation position of the target vowel. The first vowel / consonant section determination unit determines that the vowel section is a vowel section based on first articulation position calculation means and the frequency components of a plurality of single vowels whose utterance content and articulation position are known. And a second articulation position calculating means for calculating the articulation position of the target vowel whose utterance content is not calculated by the first articulation position calculation means from the frequency component of the target vowel according to a predetermined algorithm. And extracting an articulation position, which is a feature amount of the voice independent of the speaker or language, from the frequency components of the voice.

<Operation> When a voice is input, the input voice is subjected to frequency analysis to obtain a frequency component. Further, the vowel / consonant section determination unit determines whether the input voice is a vowel section or a consonant section.

Then, the first articulation position calculating means in the articulation position extraction unit determines the position on the conversion table relating to the target vowel whose utterance content determined by the vowel / consonant interval determination unit to be a vowel interval is known. Is determined based on at least two frequency components of a plurality of known single vowels and positions on the conversion table. And furthermore
Using the obtained element value at the position on the conversion table, the articulation position of the target vowel is calculated. On the other hand, the articulation position of the target vowel whose vowel content is known by the vowel / consonant interval determination unit and whose utterance content is not calculated by the first articulation position calculation means is determined by the articulation position extraction. The second articulation position calculating means in the section calculates the frequency content of the target vowel according to a predetermined algorithm based on the frequency components of a plurality of single vowels whose utterance contents and articulation positions are known.

In this way, the tonal position independent of the speaker or language is extracted from the frequency component of the voice.

<Example> Hereinafter, the present invention will be described in detail with reference to an illustrated example.

FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention. The audio signal input from the microphone 1 is amplified by the amplifier 2 and input to the acoustic analysis unit 3. In the acoustic analyzer 3, the frequency analysis of the input audio signal is performed by using a group of band-pass filters (hereinafter referred to as BPF) or by performing a fast Fourier transform on a value obtained by applying a window to the audio waveform data.

The vowel / consonant section determination unit 4 determines a vowel section and a consonant section of the audio signal. The determination of the vowel section and the consonant section is performed with reference to the power and spectrum change of the input voice. As a result, when it is determined that the vowel section is a vowel section, the articulation position of the vowel is extracted by the articulation position extraction unit 5 according to the present invention. The extraction of the articulation position of this vowel
The conversion formulas and rules for calculating the articulation position from the frequency components of the voice are read from the conversion formula storage unit 6, and the conversion is performed using the read conversion formulas and rules. On the other hand, if the vowel / consonant section determining unit 4 determines that the consonant section is a consonant section, the consonant pattern conversion unit 7 performs matching between the consonant pattern stored in the consonant pattern storage unit 8 and the frequency component of the consonant section. Thus, consonant pattern candidates are output. In this way, the input voice is converted into a time series of the vowel articulation position and the consonant pattern.

The pattern matching unit 9 includes a time series of a vowel articulation position and a consonant pattern obtained from the input voice as described above, and a standard pattern storage unit 10 obtained for each known word in advance by the same method as described above. Is calculated for each word stored in the standard pattern, and words are recognized based on the similarity calculation result. Then, the recognition result is displayed on the result display unit 11.

The present invention relates to an articulation position extraction operation of a vowel executed by the articulation position extraction unit 5, and articulation of a vowel having a known utterance content in an input word (hereinafter referred to as a vowel in a word). The position is calculated from the frequency component of the sound according to a predetermined algorithm.

Hereinafter, the first embodiment according to the present invention will be described in detail.

First Embodiment In this embodiment, as described above, based on the frequency component of a single vowel whose utterance content and articulation position are known, the articulation position of a vowel in a word is determined from the frequency component of the speech according to a predetermined algorithm. It is to be calculated. In the present embodiment, Japanese vowels (/ A
/, / A /, / c /, / d /, / o /) are used. FIG. 2 is a diagram showing articulation positions of various vowels. x represents before and after the articulation position, and the larger number is the front. Further, y represents the upper and lower positions of the articulation position, and the larger number is the lower side. In the figure,
The katakana notation enclosed by ○ is the Japanese vowel. Hereinafter, the articulation position is represented by coordinates (x, y) within the following range.

1 ≦ x, y ≦ 7 where x, y: integer values In this embodiment, it is assumed that each articulation position is on a grid point of coordinates. This can be said to be appropriate from the viewpoint of human hearing. That is, here, the articulation position of the single vowel / a / is (2,7), and the articulation position of the single vowel / a / is (6,7).
2), the articulation position of a single vowel / c / (2,2), a single vowel / d /
The articulation position of (5,4) and the articulation position of single vowel / o /
4). Then, based on the articulation position of the single vowel set in this way, the articulation position of the vowel in the word is expressed by the coordinates (x, y).

Next, the relationship between the articulation position of the vowel and the frequency component of the vowel uttered at the articulation position will be described. FIG. 3 shows the range between the first formant frequency (hereinafter, referred to as F (1)) and the second formant frequency (hereinafter, referred to as F (2)) of the Japanese vowel in FIG. 2 for each gender. FIG. FIG. 4 is a graph showing F when various speakers other than Japanese vowels in FIG. 2 are uttered by a specific speaker.
FIG. 3 is a diagram showing a relationship between (1) and F (2). That is, from FIG. 2, FIG. 3, and FIG. 4, generally, the relationship between the formant frequency and the articulation position is such that the increase / decrease of F (1) corresponds to the increase / decrease of y, and the increase / decrease of F (2) corresponds to the increase / decrease of x. It turns out that it corresponds. For some vowels (/ a / and / d /), the third formant frequency (hereinafter, F
(Represented by (3)) affects the increase and decrease of x and y. Therefore, the articulation position of the vowel in the word is estimated from these relationships and the frequency component of the vowel in the word.

Next, a method for estimating the articulation position of a vowel in a word will be specifically described.

As described with reference to FIG. 1, the input speech waveform is cut out into a vowel section or a consonant section in advance by the acoustic analysis section 3 and the vowel / consonant section determination section 4, and is subjected to acoustic analysis. The formant frequency is extracted. In the present embodiment, the formant frequency of the phoneme section labeled with the vowel as described above is used.

FIG. 5 is a flowchart of the operation of calculating the articulation position of one vowel in a word, which is executed by the articulation position extraction unit 5 of FIG.

In step S1, the formant frequency of the section determined to be a vowel section by the vowel / consonant section determination section 4 is input, and the label (that is, the utterance content) added to the formant frequency of the input vowel section is Is determined. As a result, the process proceeds to step S2, step S3, step S4, step S5, or step S6 according to the content of the determined label.

In step S2, a vowel / a / articulatory position calculation routine, which will be described in detail later, is executed, and the operation of calculating an articulation position of one vowel ends.

In step S3, a vowel / a / articulation position calculation routine, which will be described in detail later, is executed, and the operation of calculating a vowel articulation position ends.

In step S4, a vowel / c / articulation position calculation routine, which will be described in detail later, is executed, and the operation of calculating the articulation position of one vowel ends.

In step S5, a vowel / d / articulation position calculation routine, which will be described in detail later, is executed, and the operation of calculating the articulation position of one vowel ends.

In step S6, a vowel / e / articulation position calculation routine, which will be described in detail later, is executed, and the operation of calculating the articulation position of one vowel ends.

Hereinafter, the articulation position calculation routine of each vowel executed in steps S2 to S6 will be described in more detail.

(A) Vowel / a / articulatory position calculation routine In the vicinity of a single vowel / a / articulatory position, when the articulatory position changes, F (1) and F (2) change nonlinearly. Therefore, a table (hereinafter, referred to as a conversion table) for directly converting the values of F (1) and F (2) of the vowel / a / in a word to an articulation position is prepared (Table 1 shows an example thereof). (Shown) in the conversion formula storage unit 6.

This conversion table is obtained by having various speakers utter at various articulation positions and considering the relationship between the articulation positions and the formants. The coordinates of a single vowel on the conversion table (hereinafter referred to as table position) are represented by (I, J) as follows. That is,
Table position of single vowel / a / is (8,11), single vowel / d /
Is (2,4), and the single vowel / o / table position is (2,15). Here, the increasing / decreasing direction of I is F (1).
The direction of increase / decrease of F (2) indicates the direction of increase / decrease of F (2) (ie, the direction of increase / decrease of x).

Using the conversion table as described above, the articulation position of a vowel in a word is calculated from F (1) and F (2) as follows. In other words, vowels / a /
Is higher than F (2) of the single vowel / a /, the articulation position is shifted toward the articulation position of the single vowel / d /. Therefore, the vowel / a / F (1) in a word is normalized by the single vowel / a / F (1) and the single vowel / d / F (1) to obtain a table of vowels / a / in the word. I at position (I, J)
Ask for. Further, F (2) of a single vowel / a / and a single vowel /
The vowel / a / F (2) in the word is normalized by d / F (2) to determine the J of the table position (I, J) of the vowel / a / in the word. Thus, the table position (I, J) of the vowel / a / in the word is calculated. If the vowel F (2) in the word is lower than the single vowel / a / F (2), the articulation position is shifted toward the single vowel / o / articulation position. Therefore, F (1) of a vowel in a word is normalized by F (1) of a single vowel / a / and F (1) of a single vowel / o /, and the table position of the vowel / a / in the word (I , J). Further, the vowel F (2) in the word is normalized by the single vowel / a / F (2) and the single vowel / o / F (2), and the table position (I , J).
Thus, the table position (I, J) of the vowel in the word is calculated.

Then, a value on the conversion table at the calculated table position (I, J) (hereinafter referred to as TEBLE (I, J)) is obtained from the conversion table, and the following TABLE (I, J) and the articulation position (x , y), the articulation position (x, y) of the vowel / a / in the word is calculated based on TABLE (I, J) obtained from the conversion table.

[N] is the maximum integer not exceeding N. FIG. 6 is a flowchart of the articulation position calculation routine of the vowel / a / in the word in the flowchart of FIG. Here, each variable used in the description of the articulation position calculation routine of each vowel described below will be described.

F ^V (n) (V = a, i, u, e, o, n = 1, 2, 3) ... n-th formant frequency of vowel V in a word F ^V _lV (n) (V = a, i, u, e, o, n = 1,2,3) ... nth formant frequency of single vowel V (I, J) (V = a, i, u, e, o) ... table position of vowel V in word (I _V , J _V ) (V = a, j, u, e, o) ... table position of single vowel V Next, the routine for calculating the articulation position of vowel / a / in a word will be described in detail with reference to FIG. I do.

In step S11, F ^a (2) whether higher is determined than F ^a _lV (2). As a result, if F ^a (2) is higher than F ^a _lV (2), the process proceeds to step S12; otherwise, the process proceeds to step S14.

In step _{^{S12, F a lV (1)}} : F a (1): F e lV (1) = I
_a: I: to ^{_{I e, F a lV (1}} ), F a (1), F e lV (1), I a and I
I is calculated by substituting _e . F ^a _lV (2): F
_{^{a (2): F e lV}} (2) = J a: J: the ^{_{J e, F a lV (2}} ), F
_{^{a (2), F e lV}} (2), J is calculated by substituting J _a and J _e.

Here, the table position (I, J) of the above single vowel / a /
= (8,11) to I _a = 8, J _a = 11 and the table position (I, J) = (2,4) for a single vowel / d / I _e = 2, J _e = 4 .
_{^{Further, F a lV (1),}} F a lV (2), F a (1), F a (2), F e lV
The value of (1) and F ^e _lV (2), the value extracted by the acoustic analysis section 3 as described above is used.

In step S13, TABLE (I, J) is obtained according to the conversion table, based on the vowel / a / table position (I, J) in the word calculated in step S12.
Then, based on the obtained TABLE (I, J), (1)
The articulation position (x, y) is calculated from the formula, and the vowel /
The sounding position calculation routine of (a) ends.

In step _{^{S14, F a lV (1)}} : F a (1): F o lV (1) = I
_a: I: I _o I is calculated _{^{from, F a lV (2):}} F a (2): F
^o _1V (2) = J _a : J: J J is calculated from _o .

Here, the table position (I, J) of the above-mentioned single vowel / o /
= (2,15) to I _o = 2, J _o = 15.

In step S15, TABLE (I, J) is obtained according to the conversion table based on the vowel / a / table position (I, J) in the word calculated in step S14.
Then, based on the obtained TABLE (I, J), (1)
The articulation position (x, y) is calculated from the formula, and the vowel /
The sounding position calculation routine of (a) ends.

(B) Routine position calculation routine for vowel / a / The up-down direction (ie, the value of y) of the vowel / a / articulation position is determined mainly by the level of F (1). Therefore, the value of the articulation position y of the vowel / a / in the word is a single vowel /
Thresholds “BNDIE1” and “BNDIE2” determined according to the value of F (1) of a // and the value of F (1) of a single vowel / d /, and F (1) of a vowel / a / in a word Is determined according to the result of comparison with.

Further, the calculation method of the front-rear direction (that is, the value of x) of the articulation position of the single vowel / a / in the word differs depending on the magnitude of the articulation position y determined as described above. That is, when y ≦ 2, the articulation position of the vowel / a / in the word is shifted in the direction of the articulation position of the single vowel / u /. Here, as described above, since F (3) is related to the increase / decrease of the articulation position x, y in the vowel / a /, the articulation position x is calculated based on the level of F (3). That is, vowel /
The articulation position of a / is a single vowel /
When continuously changing to the articulation position of c /, F (3) decreases and approaches F (2), and overlaps or approaches F (2) (during this time, the value of F (2) is almost It does not change). The range up to this is the vowel / a / region. Furthermore, when the articulation position is behind, it enters the area of vowel / c /,
This time, F (3) does not change much, F (2) drops, and F (2) moves away from F (3) again.

Therefore, in the vowel / a / region where y ≦ 2, the value of the articulation position x of the vowel / a / in the word is:
The value of F (3) at the articulation position x = jx (= 5) when the values of F (2) and F (3) of vowel / a / overlap (hereinafter, Fjx
) And the value of F (3) of a single vowel / a // and the value of F (3) of a vowel / a / in a word.

On the other hand, when y> 2, the vowel / a /
Are shifted in the direction of the single vowel / d / articulation position. At this time, when the articulation position of the vowel / a / in the word continuously changes from the articulation position of the single vowel / a / to the articulation position of the single vowel / e /, F (2) decreases and F (2) decreases. 3)
Also decrease. Therefore, it is assumed that the change between F (2) and F (3) has a contribution ratio of 2: 1 to the change in the articulation position, and the values of F (2) and F (3) of the single vowel / a / and the single vowel / D / F
(2), the value of F (3) and the vowel / a / F of the word
Normalize the values of (2) and F (3) to obtain the vowel / a /
Is calculated.

FIG. 7 is a flow chart of a routine for calculating the articulation position of the word / a / in the words in the flow chart of FIG. Hereinafter, the articulation position calculation routine of the vowel / a / in the word will be described in detail with reference to FIG.

In step S21, it is determined whether the value of F ⁱ (1) is smaller than 250 Hz. As a result, if it is lower than 250 Hz, the process proceeds to step S22; otherwise, the process proceeds to step S23.

In step S22, the tone position y is set to y = 1, and the process proceeds to step S28.

In step S23, it is determined whether the value of F ⁱ (1) is smaller than “BNDIE1”. As a result, if it is smaller than "BNDIE1", the process proceeds to step S24; otherwise, the process proceeds to step S25.

Here, the threshold “BNDIE1” is a value set as follows.

BNDIE1 = In _{^{(5F i lV (1) +}} F e lV (1)) / 6 Step S24, articulation position y is set to y = 2, the process proceeds to step S28.

In step S25, it is determined whether the value of F ⁱ (1) is smaller than “BNDIE2”. As a result, if it is smaller than "BNDIE2", the process proceeds to step S26; otherwise, the process proceeds to step S27.

Here, the threshold “BNDIE2” is a value set as follows.

In _{^{BNDIE2 = (F i lV (1}} ) +2 e lV (1)) / 3 Step S26, articulation position y is set to y = 3, the process proceeds to step S28.

In step S27, the tone position y is set to y = 4, and the process proceeds to step S28.

In step S28, it is determined whether the value of the articulation position y set in steps S22, S24, S26, and 27 is 2 or less. as a result,
If it is 2 or less, the process proceeds to step S29; otherwise, the process proceeds to step S34.

In step S29, it is determined whether the value of F ⁱ (3) is smaller than “BNDX”. As a result, if it is smaller than “BNDX”, the process proceeds to step S30; otherwise, the process proceeds to step S30.
Continue to 31.

Here, the threshold value “BNDX” is a value set as follows.

BNDX = In _{^{(F i lV (3) +}} F jx) / 2 F jx = (F i lV (2) + F u lV (3)) / 2 step S30, articulation position x is set to x = 5, in a word The vowel / a / articulation position calculation routine ends.

In step ^{_{31, F jx: F i (}} 3): F i lV (3) = x j: x: x is calculated from the x _i.

Here, x _i = 6 from the articulation position (x, y) = (6,2) of a single vowel / a /. Further, when the articulation position of vowel / a / continuously changes from the articulation position of single vowel / a / to the articulation position of single vowel / u /, the articulation position x where F (2) and F (3) overlap. = a x _j = 5 from xj = 5.

In step S32, it is determined whether the value of the articulation position x calculated in step S31 is smaller than 6. As a result, if it is smaller than 6, the process proceeds to step S33. The sounding position calculation routine of / ends.

In step S33, the articulation position x is set to x = 6, and the articulation position calculation routine for the vowel / a / in the word ends.

In step _{^{S34, {2F i lV (2}} ) + F i lV (3)}: {2F
^{i (2) + F i (} 3)}: {2F e lV (2) + F e lV (3)} =
x is calculated from x _i : x: x _e .

Here, the articulation position (x, y) of the single vowel / a / = (6, 2) to x _i = 6, and the articulation position (x, y) of the single vowel / a /
(5,4) from a x _e = 5.

In step S35, it is determined whether or not the value of x calculated in step S34 is smaller than 4. as a result,
If it is smaller than 4, the process proceeds to step S36, otherwise, the articulation position calculation routine for the vowel / a / in the word is ended.

In step S36, the articulation position x is set to x = 4, and the articulation position calculation routine for the vowel / a / in the word ends.

(C) Routine of vowel / C / articulation position calculation routine The articulation position y of vowel / C / can be obtained by the level of F (1). That is, the vowel / u / F in the word
If (1) is higher than F (1) of a single vowel / u /, the articulation position of the vowel / u / in the word is shifted to the articulation position of the single vowel / u /, so that the single vowel / u / The value of F (1) // and the value of F (1) of a single vowel / o /
The value of (1) is normalized to calculate the articulation position y of the vowel / U / in the word. Conversely, F (1) of a vowel / u / in a word
Is lower than F (1) of a single vowel / u /, the articulation position y of the vowel / u / in the word is shifted to the y = 1 side.
The value of F (1) for a single vowel / U / and F corresponding to y = 1
By normalizing the value of F (1) of the vowel / u / in the word with the value of (1), the articulation position y of the vowel / u / in the word is calculated.

On the other hand, the articulation position x of the vowel / C / can be determined by the level of F (2). That is, if the vowel / U / F (2) in the word is higher than the single vowel / U / F (2), the articulation position of the vowel / U / in the word is a single vowel / A / articulation position. Side, the value of F (2) of a single vowel / U / and F (2) = Fj at the above articulation position x = jx (= 5)
The value of F (2) of the vowel / u / in the word is normalized with the value of x to calculate the articulation position y of the vowel / u / in the word. Conversely, a vowel / U / F (2) in a word is a single vowel / U / F
If it is lower than (2), the articulation position of the vowel / u / in the word is shifted to the articulation position side of the single vowel / u /, so the value of F (2) of the single vowel / u / and the single vowel By normalizing the value of F (2) of a vowel / U / in a word with the value of F (2) of / o /,
The articulation position x of the vowel / u / in the word is calculated.

FIG. 8 is a flowchart of a routine for calculating the articulation position of a vowel / U / in a word in the flowchart of FIG. Hereinafter, the routine for calculating the articulation position of vowels / U / in words will be described with reference to FIG.

In step S41, F ^o _lV greater than the value of the value F ^u _lV (1) (1), and, whether or not the value of F ^u (1) is greater than the value of F ^u _lV (1) is Is determined. As a result, F ^o _lV greater than the value of the value F ^u _lV (1) (1), the value F ^u _lV of F ^u (1)
If it is larger than the value of (1), the process proceeds to step S42,
Otherwise, go to step S43.

In step _{^{S42, F u lV (1)}} : F u (1): F o lV (1) = y
_u : y: y y is calculated from _o .

Here, a single vowel / U / articulatory position (x, y) = a (2,2) and y _u = 2. In addition, the articulation position of a single vowel / o / (x,
y) = (1,4) to _yo = 4.

In step _{^{S43, F u lV (1)}} = F u (1): 200 = y u: y: 1
Is calculated from

In step S44, the above steps S42 and S43
Is less than 2 and the value of ^Fu
It is determined whether the value of (1) is 300 Hz or more. As a result, if the value of the articulation position y is smaller than 2, and the value of F ^u (1) is 300 Hz or more, the process proceeds to step S45; otherwise, the process proceeds to step S46.

In step S45, the value of the tone position y is set to y = 2, and the process proceeds to step S46.

In step S46, the above steps S42 and S43
It is determined whether or not the value of the articulation position y obtained in is larger than 3. As a result, if it is larger than 3, the process proceeds to step S47; otherwise, the process proceeds to step S48.

In step S47, the value of the tone position y is set to y = 3, and the process proceeds to step S48.

In step S48, the whether or not the value of F ^u (2) is greater than the value of F ^u _lV (2) is determined. As a result, the process proceeds to step S49 if the value of F ^u (2) is greater than the value of F ^u _lV (2), the process proceeds to step S50 if not.

In step S49, F _jx : F ^u (2): F ^u _lV (2) = x _j : x: x _u
X is calculated from

Here, x _u = 2 from the articulation position (x, y) = (2,2) of a single vowel / u /. Also, from the above articulation position = x = xj = 5
x _j = 5.

In step _{^{S50, F u lV (2)}} : F u (2): F o lV (2) = x
x is calculated from _u : x: _xo .

Here, x _o = 1 from the articulation position (x, y) = (1,4) of a single vowel / o /.

In step S51, the above steps S49 and S50
It is determined whether or not the value of the articulation position x obtained in is larger than 5. As a result, if it is larger than 5, the process proceeds to step S52; otherwise, the process proceeds to step S53.

In step S52, the value of the articulation position x is set to x = 5, and the process proceeds to step S53.

In step S53, the above steps S49 and S50
The value of the articulation position x obtained in the above is x = 5, and ^Fu
It is determined whether the value of (2) is smaller than 0.9F _jx .
As a result, if x = 5 and the value of F ^u (2) is smaller than 0.9F _jx , the process proceeds to step S54. Otherwise, the articulation position calculation routine for one vowel / u / in the word is _executed. finish.

In step S54, the value of the articulation position x is set to x = 4, and the articulation position calculation routine for the vowel / u / in the word ends.

(D) Routine of vowel / d / articulation position calculation routine
Is different from the case where the tone position of the single vowel / a / is shifted to the position of the tone position of the single vowel / a /. That is, F (1) of a vowel / d / in a word
Is greater than the threshold "BNDIE2" of F (1) for defining the boundary between the articulation positions y = 3 and y = 4, the articulation position of the vowel / e / in the word is a single vowel / a / It is determined that it is shifted to the articulation position side. Then, as described in (A), the single vowel / a / F (1) and the single vowel / d / F
(1) normalizes F (1) of the vowel / d / in the word with
Table of vowels / d / in words by normalizing vowels / d / F (2) in words with single vowels / a / F (2) and single vowels / d / F (2) Calculate the position (I, J). Then, TABLE (I, J) is obtained from the conversion table based on the table position (I, J), and the articulation position (x, y) of the vowel / d / in the word is calculated from the equation (1). is there.

Conversely, the value of F (1) of the vowel / d / in the word is the threshold "BN
If it is smaller than DIE2 ", it is determined that the articulation position of the vowel / e / in the word is shifted to the articulation position of the single vowel / a /. In that case, the vowel in the word described in (B) is used. By using the same algorithm as the articulation position calculation algorithm (FIG. 7) when the articulation position of / a / is shifted to the articulation position side of the single vowel / e /, the articulation position of the vowel / e / in the word is calculated. However, in this case, the step of setting the value of the articulation position y to y = 1 and the step of calculating x when y = 1 are excluded.

FIG. 9 is a flowchart of a routine for calculating the articulation position of a vowel / d / in a word in the flowchart of FIG. Hereinafter, the articulation position calculation routine of the vowel / d / in the word will be described with reference to FIG.

In step S61, it is determined whether the value of F ^e (1) is greater than the value of the threshold “BNDIE2”. As a result, "BNDI
If it is larger than the value of E2 ", the process proceeds to step S62; otherwise, the process proceeds to step S64.

In step _{^{S62, F a lV (1)}} : F e (1): F e lV (1):
I is calculated from I _a : I: I _e and F ^a _lV (2): F ^e (2): F ^e _{lV
(2) = J a: J} : J is calculated from the J _e.

In step S63, TABLE (I, J) is obtained according to the conversion table based on the vowel / d / table position (I, J) in the word calculated in step S62.
Then, based on the obtained TABLE (I, J), (1)
The articulation position (x, y) is calculated from the formula, and the vowel /
The sounding position calculation routine of d / ends.

In step S64, it is determined whether the value of F ^e (1) is smaller than “BNDIE1”. As a result, if it is smaller than "BNDIE1", the process proceeds to step S65; otherwise, the process proceeds to step S66.

In step S65, the value of the tone position y is set to y = 2, and the process proceeds to step S69.

In step S66, it is determined whether the value of F ^e (1) is smaller than “BNDIE2”. As a result, if it is smaller than "BNDIE2", the process proceeds to step S67; otherwise, the process proceeds to step S68.

In step S67, the value of the tone position y is set to y = 3, and the process proceeds to step S69.

In step S68, the value of the tone position y is set to y = 4, and the process proceeds to step S69.

In step S69, {2F ⁱ _lV (2) + F ⁱ _lV (3)}: {2F
^e (2) + F ^e (3)｝: {2F ^e _lV (2) + F ^e _lV (3)} =
x is calculated from x _i : x: x _e, and the articulation position calculation routine for the vowel / d / in the word ends.

(E) The vowel / o / articulation position calculation routine is performed by the vowel / o / articulation position calculation routine.
Is different from the case where the tone position of the single vowel / a / is shifted to the position of the tone position of the single vowel / a /. That is, F (1) of a vowel / o / in a word
Is greater than the F (1) threshold "BNDOU2" for defining the boundary between the articulation positions y = 3 and y = 4, the articulation position of the vowel / d / in the word is a single vowel / a / It is determined that it is shifted to the articulation position side. Then, as described in (A), the single vowel / a / F (1) and the single vowel / o / F
(1) normalizes the vowel / o / F (1) in a word while the single vowel / a / F (2) and the single vowel / o / F
By normalizing F (2) of the vowel / o / in the word with (2), the table position (I,
J) is calculated. Then, TABLE (I, J) is obtained from the conversion table based on the table position (I, J),
The articulation position (x, y) of the vowel / o / in the word is calculated from the equation (1).

Conversely, the value of F (1) of the vowel / o / in the word is equal to the threshold "BN
If it is smaller than "DOU2", it is determined that the articulation position of the vowel / o / in the word is shifted toward the articulation position of the single vowel / u /. In that case, the vowel in the word described in (C) is determined. The articulation position of the vowel / o / in the word is calculated by the same algorithm as the articulation position calculation algorithm (FIG. 8) when the articulation position of / u / is shifted to the articulation position side of the single vowel / o /. However, in this case, the value of the tone position y is set based on the thresholds “BNDOU1” and “BNDOU2”.

FIG. 10 is a flowchart of a routine for calculating an articulation position of a vowel / o / in a word in the flowchart of FIG. Hereinafter, the articulation position calculation routine of the vowel / o / in the word will be described with reference to FIG.

In step S71, it is determined whether the value of F ^o (1) is greater than the value of the threshold “BNDOU2”. As a result, "BNDO
If it is larger than the value of U2 ", the process proceeds to step S72; otherwise, the process proceeds to step S74.

Here, the above “BNDOU2” is a value set as follows.

BNDOU2 = In _{^{(F u lV (1) +}} 2F o lV (1)) / 3 Step _{^{S72, F a lV (1)}} : F o (1): F o lV (1) = I
_a: I: I _o I is calculated _{^{from, F a lV (2):}} F o (2): F
^o _1V (2) = J _a : J: J J is calculated from _o .

In step S73, TABLE (I, J) is obtained according to the conversion table based on the vowel / o / table position (I, J) in the word calculated in step S72.
Then, based on the obtained TABLE (I, J), (1)
The articulation position (x, y) is calculated from the formula, and the vowel /
The sounding position calculation routine of (e) ends.

In step S74, it is determined whether the value of F ^o (1) is smaller than “BNDOU1”. As a result, if it is smaller than "BNDOU1", the process proceeds to step S75; otherwise, the process proceeds to step S76.

Here, the above “BNDOU1” is a value set as follows.

BNDOU1 = In _{^{(5F u lV (1) +}} F o lV (1)) / 6 step S75, the value of the articulation position y is set to y = 2, the process proceeds to step S79.

In step S76, it is determined whether the value of F ^o (1) is smaller than “BNDOU2”. As a result, if it is smaller than "BNDOU2", the flow proceeds to step S77; otherwise, the flow proceeds to step S78.

In step S77, the value of the tone position y is set to y = 3, and the process proceeds to step S79.

In step S78, the value of the tone position y is set to y = 4, and the process proceeds to step S79.

In step _{^{S79, F u lV (2)}} : F u (2): F o lV (2) = x
x is calculated from _u : x: _xo, and the articulation position calculation routine for the vowel / o / in the word ends.

Vowel articulatory position calculation algorithms and formant frequency of each vowel in each word, as described above, articulation position of each single vowel (x, y), the table position of each vowel (I _V, J _V), the conversion The table, each threshold value, and the like are stored in the conversion formula storage unit 6 shown in FIG. 1. When the articulation position extraction unit 5 performs the operation of calculating the articulation position of a vowel in a word, the conversion formula storage unit 6
Is read from. Here, in the present embodiment, the first articulation position calculating means performs steps S11 to S11 in FIG.
15 and correspond to steps S61 to S63 in FIG. 9 and steps S71 to S73 in FIG. On the other hand, the second articulation position calculating means includes steps S21 to S36 in FIG. 7, steps S41 to S54 in FIG. 8, steps S64 to S69 in FIG.
This corresponds to steps S74 to S79 in FIG.

As described above, in the present embodiment, the articulation position of a vowel in a word is calculated from the formant frequency of a vowel in a word of an input voice using a known formant frequency of a single vowel according to a predetermined algorithm. Like that. Therefore, according to the present embodiment, it is possible to extract an articulation position, which is a feature amount of speech independent of a speaker or language, based on a formant frequency characterizing a phoneme.

It goes without saying that the conversion table in the present embodiment is not limited to the one illustrated in Table 1.

 Next, a second embodiment will be described in detail.

Second Embodiment In this embodiment, a rule for generating the articulation position of a vowel from the frequency component of the vowel in the word is automatically created by learning using the neural network,
According to this rule, the articulation position of a vowel in a word is generated.

Here, an outline of the neural network will be described. A neural network is a network that mimics the structure of the human brain, and is formed by connecting a plurality of units corresponding to neurons in the brain in a complex manner.

The structure of the above unit is composed of a part that receives an input signal from another unit, a part that converts the input signal according to a certain rule, and a part that outputs the result of the conversion. The plurality of units form a hierarchical network composed of an input layer, an intermediate layer, and an output layer, as will be described in detail later. Coupling coefficients indicating the strength of coupling are attached to coupling portions with other units. ing.

The coupling coefficient indicates the strength of coupling between the units, and changing the value of the coupling coefficient changes the network structure. That is, the learning of the neural network means that data belonging to one of two events having a certain known relationship is sequentially input to the input layer of the network formed in the hierarchical structure, and at that time, the output is Changing the coupling coefficient so as to reduce the difference between output data output to the layer and data belonging to another event corresponding to the input data (ie, target value = teacher data). . In other words, the structure of the network is changed so that when data belonging to one of two events having a predetermined relationship is input, data corresponding to the other event is output.

The neural network used in this embodiment has a structure as shown in FIG. In other words, this neural network consists of input layers 2 in order from the bottom in the figure.
1. It has a three-layer structure consisting of an intermediate layer 22 and an output layer 23.
The input layer 21 has 16 units 24, 24,.
, And 10 units 25, 25,..., And the output layer 23 has seven units 26a, 26b,..., 26g and seven units 27a, 27b,.
There are 14 units consisting of 7g.

Here, one of the 16 units 24 of the input layer 21 includes 16 units.
An output signal is input from one channel of the channel BPF group. Also, one of the seven units 26 of the output layer 23
26g are coordinate values of the articulation position x (1 to 1 shown in FIG. 2).
7) (for example, it is assumed that the coordinates 1 → 7 of x correspond in the order of the units 26a → 26g) and the other 7
, 27g are coordinate values of the articulation position y (second
Any one of 1 to 7 shown in the figure (for example, unit 27a → 2
(It is assumed that the coordinates 1 → 7 of y correspond to the order of 7g). Each of the units 24, 24,... Of the input layer 21 is connected to all the units 25,. .. Of the intermediate layer 22 are connected to all the units 26a,..., 26g, 27a,. However, the units in each layer are not connected.

The neural network having the above structure is stored in the conversion formula storage 6 together with the coupling coefficient.

The neural network having the above structure operates as follows.

The frequency components of the vowels in the input voice are input to the units 24, 24,... Of the input layer 21. That is, in this embodiment, the output values from the 16-channel BPF group are input to the corresponding units 24,..., 24 of the input layer 21 for each channel. The center frequency of this BPF group is 300Hz ~ 3400
The value obtained by dividing the frequency of Hz into 16 equal intervals on the mel scale is used.

Then, the input output values from the BPF group are converted by the sigmoid function in each unit 24, 24, and are transmitted to each unit 25, 25, in the intermediate layer 22. At that time, each unit 25, 25,.
, The sum of values obtained by multiplying the output values of the units 24, 24,... Of the input layer 21 by the above coupling coefficient is input. Similarly,
Each unit 25, 25,... Of the intermediate layer 22 is a unit of the input layer 21.
The values input from 24, 24,... Are converted by the sigmoid function and output to the units 26a,..., 26g, 27a,. Each unit 26a, ..., 26g, 27a, ..., 27 of the output layer 23
The sum of the values obtained by multiplying the output values of the units 25, 25,... of the intermediate layer 22 by the coupling coefficient is input to g. , 26g, 27a,..., 27g of the output layer 23 convert the values input from the units 25, 25,.

Here, the sigmoid function f (x) is given by the following equation.

f (x) = 1 / (1 + exp (−x + a)) a: constant As described above, each unit 26a,..., 26g, 27 of the output layer 23
From the output values output from a,..., 27g, the articulation position (x, y) is determined as follows. That is, of the seven units 26a,..., 26g corresponding to the articulation position x of the output layer 23, the coordinate value of x corresponding to the unit outputting the largest value (for example, the unit 26b) ( For example, x = 2)
Is the value of the articulation position x. Further, of the seven units 27a,..., 27g corresponding to the articulation position y, y corresponding to the unit outputting the largest value (for example, 27g)
(For example, y = 7) is set as the value of the articulation position y.
Thus, the articulation position (x, y) (for example, (2, 7)) is determined.

Learning of the neural network is performed as follows using an error back propagation algorithm as a learning algorithm. That is, first, the units 24, 2 of the input layer 21
Output values from the BPF group (for example, target value (2,7)
Is output.) Next, among the units 26a, 26b,..., 26g that output the coordinate values of the articulation position x of the output layer 23, the unit (for example, the unit 26b) corresponding to the target value (for example, x = 2) of the articulation position x Enter “1” only for other units (for example, units 26a, 26c, ..., 26g)
Input "0". On the other hand, among the units 27a, 27b,..., 27g that output the coordinate values of the articulation position y of the output layer 23, the units (for example, the unit 27g) corresponding to the target value (for example, y = 7) of the articulation position y Only “1” is input, and “0” is input to other units (for example, units 26a, 26b,..., 26f). Then, the amount of change of the coupling coefficient with respect to the target value (for example, (2, 7)) is obtained by the error back propagation algorithm, and a new coupling coefficient between the units is set. When the above operation is repeated several times, eventually, the units 26a, 26b, which output the coordinate value of the articulation position x of the output layer 23,
, 26g, only the unit (eg, unit 26b) corresponding to the target value (eg, x = 2) outputs “1” and the other units output “0”, while the other unit outputs “0”. Among the units 27a, 27b,..., 27g that output the coordinate values of the articulation position y, only the unit (for example, the unit 27g) corresponding to the target value (for example, y = 7) outputs “1” and outputs the other. The coupling coefficients between the input layer 21 and the intermediate layer 22 and between the intermediate layer 22 and the output layer 23 are set so that the unit outputs “0”.

The learning as described above is performed for various speakers and languages. Then, finally, the value of the coupling coefficient converges, and a correct articulation phase (x, y) is output with respect to the input of the frequency component of the vowel in most words. That is, in the neural network that has been sufficiently trained as described above, a rule for directly generating an articulation position from the output value (frequency component of a vowel in a word) of each BPF group is automatically created. I have. Therefore, the trained neural network can output the correct articulation position (x, y) for all input values.

As described above, in the present embodiment, the value of the articulation position of a vowel in a word can be generated from the output value (ie, frequency component) of the BPF group in the vowel in the word. Therefore, according to the present embodiment, it is possible to directly extract an articulation position, which is a feature amount of a sound independent of a speaker or a language, based on a frequency component of the sound.

In the second embodiment, one unit for inputting the pitch value of the audio waveform to the input layer 21 may be added, and the number of units of the input layer 21 may be set to 17. In this case, the value of the pitch may be obtained by using a conventional technique such as an autocorrelation method or a cepstrum method. The average and standard deviation of the male pitch are 125 Hz and 20.5 Hz, respectively.
The average and standard deviation of the female pitch is about 2 times the male value, respectively.
Equal to twice. In general, female frequency components are shifted to higher frequencies than male frequency components.
This is due to physiological differences between men and women.
Therefore, by inputting the information of the pitch value into the input layer 21, the male frequency component and the female frequency component are normalized by the neural network, and the articulation position of the vowel in the word can be calculated more accurately. Can be expected.

In the second embodiment, the units 24, 24,
… Input the output value from the BPF group. However, the present invention is not limited to this, and the formant frequency used in the first embodiment may be input, and the tone position may be obtained from the formant frequency.

In the second embodiment, the input layer 21 or the intermediate layer
A sigmoid function is used as a conversion function when converting the input signal input to 22. However, the present invention is not limited to this, and a threshold function may be used.

In the second embodiment, the number of units of the input layer 21, the intermediate layer 22, and the output layer 23 constituting the neural network depends on the number of channels and the articulation position of the BPF group for outputting the frequency components to be input to the neural network. Needless to say, it may be appropriately changed according to the number of x coordinate values and y coordinate values to be represented.

<Effects of the Invention> As is apparent from the above description, the speech feature extraction device of the present invention includes the vowel / consonant interval determination unit, the conversion table, and the articulation position extraction unit, and the first of the upper articulation position extraction units. Using the conversion table, the articulatory position calculation means obtains a position on the conversion table relating to the target vowel whose utterance content determined to be a vowel section by the vowel / consonant section determination unit is known. While the articulation position of the target vowel is calculated using the element value at the position on the obtained conversion table, the second articulation position calculation means calculates a plurality of single vowels whose utterance content and articulation position are known. Based on the frequency component, the vowel / consonant interval determination unit determines that the vowel is a vowel interval, and the utterance content whose articulation position is not calculated by the first articulation position calculation unit is known. Since the articulation position of the target vowel is calculated from the frequency component of the object vowel, it is possible to generate an articulation position, which is a feature amount of speech independent of a speaker or a language, from the frequency component of the vowel.

Therefore, according to the present invention, it is possible to accurately extract a characteristic amount unique to the utterance of a voice by a simple process.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of a speech recognition apparatus according to the present invention, FIG. 2 is a diagram showing articulation positions of various vowels, and FIG. 3 is a first formant frequency and a second formant frequency in Japanese vowels. FIG. 4 is a diagram showing a relationship between a first formant frequency and a second formant frequency in various vowels of a speaker, and FIG. 5 is an articulation position calculating operation of one vowel in a word. FIG. 6 is a flowchart of a routine for calculating an articulation position of a vowel / a / in a word in FIG. 5, FIG. 7 is a flowchart of an articulation position calculation routine of a vowel / a / in a word in FIG. FIG. 8 shows vowels / words in FIG.
9 is a flowchart of a routine for calculating the articulation position of /, FIG. 9 is a flowchart of a routine for calculating the articulation position of a vowel / d / in a word in FIG. 5, and FIG. 10 is an articulation of a vowel / o / in a word in FIG. FIG. 11 is a flowchart of the position calculation routine, and FIG. 11 is an explanatory diagram of the production of the neural network. DESCRIPTION OF SYMBOLS 1 ... Microphone 2 ... Amplifier 3 ... Sound analysis part 4 ... Vowel / consonant section determination part 5 ... Articulation position extraction part 6 ... Conversion type storage part 7 ... Consonant pattern conversion part , 8 ... consonant pattern storage unit, 9 ... pattern matching unit, 10 ... standard pattern storage unit, 11 ... result display unit, 21 ...
Input layer, 22 ... Intermediate layer, 23 ... Output layer, 24 ... Input layer unit, 25 ... Intermediate layer unit, 26a-26g ... Unit for outputting articulation position x, 27a-27g ... Articulation A unit that outputs the position y.

Claims (1)

(57) [Claims] A vowel / consonant section determining unit for determining a vowel section and a consonant section of an input voice in a voice feature extracting apparatus for frequency-analyzing an input voice and extracting a voice feature from an obtained frequency component; A table for converting at least two frequency components of a vowel into an articulation position, the conversion table including element values of a plurality of single vowels whose utterance contents are known, and the articulation of a target vowel whose utterance content is known A vowel / consonant interval determination unit that determines the position of the vowel in the conversion table according to the target vowel whose known vowel content is known. Is calculated based on the two frequency components of the single vowel and the position on the conversion table, and using the obtained element value at the position on the conversion table. A first articulation position calculating means for calculating an articulation position of the target vowel; and a vowel section by the vowel / consonant section determination unit based on frequency content of a plurality of single vowels whose utterance content and articulation position are known. And a second articulation position for calculating the articulation position of the target vowel whose utterance content is not calculated by the first articulation position calculation means from the frequency component of the object vowel according to a predetermined algorithm. An audio feature extraction device, comprising: calculation means for extracting, from frequency components of the audio, an articulation position which is a feature amount of audio independent of a speaker or language.