JPH03111900A - Voice recognition system - Google Patents

Abstract

PURPOSE:To enable high word recognition and to perform more difficult single sound recognition at a high recognition rate by extracting features of a feature pattern and a standard pattern while separating them by directions and performing similarity calculation in units in a time-base direction. CONSTITUTION:A voice is inputted to an analysis part 2 through a voice input part 1 to find a time space pattern, which is sectioned by a word section segmentation part 3 into words in recognition units and supplied to a feature extraction part 4. The output of the segmentation part 3 is inputted to a normalization part 41 to normalize the time base linearly. The normalized time space pattern is supplied to a spectrum field extraction part 42 to extracts a spectrum field. A pattern generation part 43 generates two-dimensional patterns by directions from the spectrum field pattern and a gradation processing part 5 performs time space gradation processing to improve the voice recognition rate. Features extracted as to various words are stored as standard pattern in a storage part 6. A calculation part 7 calculate the similarity to the standard pattern in the feature pattern storage part 6 from the processing part 5 and outputs the most similar pattern as a recognition result.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a speech recognition method, and more specifically, a standard pattern obtained using a vector field pattern and a characteristic pattern corresponding to speech to be recognized. This is related to calculating the degree of similarity with.

[Prior art]

Generally, in speech recognition, a standard pattern of speech obtained by extracting features from the word to be recognized is prepared for each word, and feature patterns extracted in the same way from the speech input as recognition target and multiple standards are prepared for each word. A method is adopted in which a standard pattern with the highest similarity is found by matching the patterns, and it is determined that a word related to this standard pattern has been input. Conventionally, the spatio-temporal pattern itself of a scalar field with the horizontal axis as the time axis and the vertical axis as the spatial axis, which is obtained by analyzing the audio signal, has been used as the feature pattern.

A typical spatio-temporal pattern of such a scalar field is a spectrum with frequency as its spatial axis, as well as a cepstrum with quefrency as its spatial axis, and a PA pattern with quefrency as its spatial axis.
Various spatiotemporal patterns such as RCOR coefficient, LSP coefficient, and vocal tract cross-sectional area function were used.

In addition, one of the issues to be solved in the field of speech recognition is dealing with multiple speakers or unspecified speakers, and improving the recognition rate by preparing a large number of standard patterns for one word. I was planning. Furthermore, even if speakers are the same, their pronunciation speeds may differ, and a DP matching method that can absorb time axis fluctuations has been developed to cope with such cases.

Conventional methods that use the spatio-temporal pattern itself in a scalar field do not necessarily achieve a sufficient recognition rate when targeting a large number of speakers or unspecified speakers. Even if we prepared a large number of standard patterns or used the leaf matching method, these problems could not be fully resolved.

Therefore, the practical application of speech recognition systems for large speech speakers or unspecified speakers has stalled.

Therefore, one of the inventors of the present invention obtained a spectral vector field pattern by spatially differentiating the spectrum of a scalar field, which is a time-frequency spatiotemporal pattern, in Japanese Patent Application Laid-Open No. 60-59394. We proposed a method to use it as a feature. Taking this a step further, the present inventors have developed a speech feature extraction method and speech recognition method that is suitable for syllable recognition and word recognition and that can obtain a high recognition rate. 62-136377
proposed by No.

The basic characteristics of this speech recognition method are to obtain a spatio-temporal pattern of a scalar field defined by a time axis and a spatial axis from an audio signal, and to convert the spatio-temporal pattern into a spatial machine.

By dividing, each grid point in the space is converted into a vector field pattern having a magnitude and direction, and for the vector of the vector field pattern, its direction parameter is quantized into N pieces (N: an integer), and this quantized value is Separate the same vectors and create N two-dimensional patterns for each direction with the size of each vector as the value of each grid point, and create a standard created in advance using the two-dimensional patterns for each direction. The point is that the input speech is identified by calculating the degree of similarity between the pattern and the feature pattern, which is the direction-specific two-dimensional pattern obtained from the speech signal input as a recognition target.

[Problem to be solved by the invention]

Although the above-mentioned speech recognition method can obtain a high word recognition rate, the challenge has been to perform the more difficult monosyllable recognition at a high recognition rate.

The present invention has been made with the aim of achieving the above object.

[Means to solve the problem]

The speech recognition method according to the present invention obtains a spatiotemporal pattern of a scalar field defined by a temporal axis and a spatial axis from an audio signal,
By spatially differentiating the spatio-temporal pattern, it is converted into a vector field pattern with magnitude and direction at each grid point in the space, and the direction parameters of the vectors of the vector field pattern are quantized into N pieces (N: an integer). , separate this quantized value into vectors with the same value, create N two-dimensional patterns for each direction with the size of the vector as the value of each grid point, and create in advance using the two-dimensional patterns for each direction. In a speech recognition method that identifies an input speech by calculating the similarity between a standard pattern that has been set and a feature pattern that is a two-dimensional pattern according to direction obtained from a speech signal input as a recognition target, standard patterns and The method is characterized in that the degree of similarity is calculated for each of the directions of each feature pattern.

Further, the present invention is characterized in that this calculation for each direction is performed for each unit in the time axis direction of the two-dimensional pattern for each direction.

[Effect]

Feature extraction is performed separately for each direction. Therefore, similar changes in the spectrum are extracted in the spectral vector field for each direction. Therefore, if the similarity is calculated for each direction, the similarity between the standard pattern and the characteristic pattern can be determined more accurately. Furthermore, when the similarity is calculated for each unit in the time axis direction, the influence of normalization distortion performed during feature extraction processing is reduced, and thereby accurate similarity can be obtained.

〔Example〕

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below based on drawings showing embodiments thereof.

FIG. 1 is a block diagram showing the configuration of an apparatus for implementing the method of the present invention. In this embodiment, the analysis section spectrally analyzes the audio signal and uses a spectrum with the frequency axis as the spatial axis as the spatiotemporal pattern of the scalar field.

Input of voice for standard pattern creation or recognition target voice is performed using a voice detector such as a microphone and A/D&
The audio signal obtained by this is analyzed using a plurality of bandpass filters (for example, 10 to 30) connected in parallel, each having a different pass frequency band. The data is input to section 2. The analysis section obtains a spatio-temporal pattern as a result of the analysis, and this pattern is segmented by the word section cut-out section 3 into each recognition unit word and provided to the feature extraction section 4 . As the word section cutting section 3, a conventionally known one may be used.

In the following description, a group of bandpass filters will be used as described above as an analysis unit that divides the audio signal into frequency bands, but a fast Fourier transformer may also be used.

The input pattern to the feature extraction unit 4 is a spatiotemporal pattern with the horizontal axis as the time axis and the vertical axis as the frequency.
The spatiotemporal pattern shown in Figure 2 extracted by
(t, x) (where t is a number indicating the sampling time, and X is a number specifying the channel number or frequency band of the bandpass filter.

1≦t≦T, 1≦X≦L).

The output of the word section extraction section 3 is input to the normalization section 41 of the feature extraction section 4, and the normalization section 41 linearly normalizes the time axis.

This is to absorb to some extent the length and shortness of the words and the length of the human voice, and the time axis is changed from T frames to M frames (for example, about 16 to 32 frames). Specifically, M≦
In the case of T, the normalized spatio-temporal pattern F(t, x) is obtained by (1) 2 below.

However, 1≦L≦M And if M>T, then F (t, x) = f l, Bye.

Figure 3 shows the spatiotemporal pattern F (
t, x).

Note that the above example is a case of linear normalization, but when performing nonlinear normalization, for example, the spectral vector field of f (t, x) is obtained in the same manner as described below, and this vector field density is It is sufficient to use hector field density equalization, etc., which is made constant.

A spectral vector field is extracted from the normalized spatio-temporal pattern in the spectral field extraction section 42 as described below. This spectral vector field is calculated using eight neighboring values of each grid point (t, x) of the normalized spatio-temporal pattern as shown in Table 1.

Table 1 X = F (t+1.x+1) +2F(t+1.
x) + F (t+1.x-1)-F (t-1,
x+1) -2P(t-1, x) -F (t-1,
x-1) Y = F (t-1, x+1) +2F(
t, x+1) + F (t+1.x+1)F(t-
1, x-1) -2P(t, x-1) -F(t+1
．． x-1) r = J X2+Y" - (5)...
・(3) ...(4) θ=jan-' ...(6) As 3 (r, θ) is the spectral vector field pattern.

To explain a little about equations (3) to (6), X is a value obtained by weighting the increment in the time axis direction of the target data in the frequency axis direction, and is the differential value in the time axis direction, that is, in the time axis direction. It can be said to be a change indicator.

Similarly, Y can be said to be a differential value in the frequency axis direction, that is, a change index in the frequency axis direction.

In a vector field in which these two indices are orthogonal coordinates, r represents the magnitude of the vector, and θ represents the direction of the vector.

This spectral vector field pattern S(r.

θ), a direction-based two-dimensional pattern creation unit 43 creates a direction-based two-dimensional pattern. That is, first, the direction parameter θ of the vectors of all the lattice points of the vector field pattern is quantized to N values. FIG. 4 is a diagram for explaining an example of quantization when N=8, and as shown in Table 2, θ and N correspond.

(Margin below) Table Next, from among the vectors of all the grid points, vectors with the same N value, which is the direction quantization value, are separated and taken out for each N value, and the magnitude of each vector is calculated. N direction-specific two-dimensional patterns 7 H (1+L θ) are created using the values of the grid points. FIG. 5 is a schematic diagram of this directional two-dimensional pattern, in which r exists only at positions corresponding to the value of N, and is 0 at other positions. As understood from equations (3) and (4), 8 neighbors are required to calculate X and Y, so
The calculated S(r, θ) is t=1.
t=M columns, and x=1. It is not calculated for the row where x=L.

Therefore, in this directional two-dimensional pattern H (t, x, θ), there are M-2 columns in the time axis direction and L-2 rows in the frequency axis direction.

Note that the value of N is not limited to 8.

Now, the two-dimensional pattern H(t
, χ, θ) is subjected to blurring processing by the blurring processing unit 5. Blur processing is performed by multiplying the pattern to be processed by 9 neighboring mask patterns weighted according to its position. If the two-dimensional pattern by direction after blurring processing is H (t, x, θ), ...(7) It can be expressed as.

Here, ωj (J·θ~8) is a mask pattern for blurring processing, and has, for example, the following values (8) and (9), and ω at the center. is at the position of the data to be processed, and ω1
~ω6 corresponds to data at eight neighboring positions.

time time Also (α.

β, ) j・θ〜8 is as shown in Table 3. It is.

(Margins below) Table 3 This (αj, βJ) is ω. The position of the target data to which tuna~ω8 is made to correspond, and the position of 8 neighboring data to which tuna to ω8 are made to correspond are specified.

Equations (8) and (9) mean that the blurring process in the time axis direction is performed more aggressively than the blurring process in the frequency axis direction.

When used to extract the features of voices of only one gender (male or female), blurring processing in the frequency axis direction is not performed as in equation (8), and the characteristics of voices of both genders (male and female) are not performed. When performing extraction, some blurring processing in the frequency axis direction is also performed as shown in equation (9).

The features extracted by performing the pokashi processing are those in which the fluctuations of the characteristics unique to the voice used for extraction are reduced. In other words, it is possible to stabilize spatiotemporal variations in features caused by different speakers or different rates of occurrence.

Therefore, if this is used as a standard pattern or an unrecognized pattern, the speech recognition rate can be improved.

However, the reason why blurring processing in the time axis direction is actively performed is because the time axis is related to the speaking speed, which varies greatly depending on the time of speech and the speaker, and the purpose is to eliminate the negative effects of this variation.

Furthermore, in the case of voices of both genders, the frequency distribution differs between men and women, and the adverse effects of this variation are eliminated by blurring the frequency axis at the same time.

Note that it is best to repeat the blurring process multiple times according to equation (7), but in the case of only one sex, repeat the blurring process with a blurring effect only on the time axis about 4 to 7 times as shown in equation (8). , in the case of both genders, as shown in (9) 2, the spatiotemporal blurring process is performed about 4 times each, with the weight of the spatial axis being about 174 to 1/8 of the weight of blurring to the time axis. It is appropriate to repeat the process.

The features extracted in advance for various words in this manner are stored as standard patterns in the standard pattern storage section 6 together with data specifying them. Then, during recognition, the calculation unit 7 calculates the degree of similarity between the characteristic pattern output from the blurring processing unit 5 obtained for the voice input as the recognition target and the standard pattern in the standard pattern storage unit 6. , data identifying the most similar standard pattern is output as a recognition result.

The first method of the present invention is characterized by the following similarity calculation. To summarize, first, for each of the N patterns shown in FIG. 5, the degree of similarity between the characteristic pattern and the standard pattern is calculated. This calculation includes the city distance Dc and the Euclidean distance 0. The correlation coefficient C is used as an index for determining the degree of similarity. Although not limited to this, when using the correlation coefficient C, the highest recognition rate can be obtained. Furthermore, in the second method of the present invention, similar calculations are performed for each N direction and also for each time axis unit, that is, for each frame.

Next, DC, D, and C will be explained.

City streets! IDc is the sum of absolute values of differences at the same position between patterns to be compared, and has the advantage of being easy to calculate.

Euclidean distance refers to the mathematically exact distance between the patterns being compared.

The correlation coefficient C is an index representing the R4Q degree of the patterns to be compared.

Z the standard pattern! (un X* θ), and the feature pattern of the speech to be recognized is 1 (t,

D,=Σ4Σ(1(t,X,θ)−Z!(t
,X.

@LIK θ))2 Further, the calculation for each frame according to the second method can be expressed as follows.

However, T' is the total number of frames (-M-2) and T' xX
When calculation is performed as an xN-dimensional vector (hereinafter referred to as the first comparison method), it can be expressed as follows. (However, X is the total number of X (-L-2)) In this case, the city distance can also be calculated as
) Furthermore, when calculation is performed for each T' number of frames (hereinafter referred to as the second comparison method), it can be expressed as follows.

Next, an experiment conducted to compare the advantages and disadvantages of these four methods will be explained. The experiment was conducted on 101 discretely generated Japanese monosyllables.

The vocalizations were made by an adult male and were normal.

The average vocal length is 440 m5ec.

A/D conversion in the audio input section 1 has an accuracy of 12 bits,
Further, the sampling frequency was 12.5 kHz.

The filter in the analysis section 2 is a 20-channel bandpass filter. Also, the frame interval is 5.12n+s
It is ec. The two-dimensional speech pattern obtained by normalizing according to these specifications is 32 (=T') xlB (=X) x8 (=N
).

Experiment 1 One standard pattern was created using data from 10 people, and the speech uttered by the same 10 people was recognized.
lose) Speaker experiment n An open speaker experiment in which one standard pattern is created using data from 29 people, and the voice uttered by another person is recognized ■ uttered by one person A standard pattern is created using data from 10 times, and the data from the same 10 times is recognized.Experiment using closed data of a specific speaker (however, for 2 speakers) Table 4 shows the results of the above-mentioned similarity calculations for these experiments. .

(Left below) According to the experimental results of Mugi and above, the urban area distance Dc using the same calculation formula
In all of Experiments 1 to N, except for 1. of the present invention.
The second method is the first. A higher recognition rate was obtained than the second comparative method, and a higher recognition rate was obtained than either the second method or the first method except for the correlation coefficient C in experiment (2).

Furthermore, the first aspect of the present invention. In the second method, the correlation coefficient C has a higher recognition rate than the Euclidean distance.

Japanese Patent Application No. 62-248915, which is the basis of the method of the present invention.

The voice recognition method of No. 62-136377 is No. 1. As shown in the results of the second comparison method, a high recognition rate can be obtained even for monosyllables, but the first comparison method of the present invention achieves a high recognition rate even for monosyllables. According to the second method, it is possible to achieve a -higher recognition rate.

The reason why a high recognition rate can be obtained by performing the associativity calculation for each direction as in the present invention is considered to be because the spectral vector field is direction-specific as described above. Furthermore, the reason why a high recognition rate is obtained when calculation is performed for each frame is considered to be because the influence of the blurring process described above can be avoided.

In other words, it is considered that frame-by-frame processing is less affected by denormalization caused by blurring processing.

Furthermore, the fact that a high recognition rate can be obtained using the correlation coefficient C is considered to be an effect unique to monosyllable recognition.

That is, when speech recognition is generally performed using spectra, the loudness of the sound is normalized for each frame at the stage of obtaining a spectral pattern. This removes the DC component. In the above experiment, this was done using the following calculations.

1logA per frame, +logA2 += +logA
, ) /20=88 (however, ^1~A, . is the original output of each of the 20 channel filters), B+=IogA+ AAB2=IogA
2AABzo=10gAzo-8A are the outputs of the filters of 20 channels in one frame. That is, the spatiotemporal pattern f (t, x) of the spectrum in FIG. 2 is a collection of t frames of Bt-Bzo.

If this process is performed, substantially the same result will be obtained regardless of whether DC, D, or C is used for speech recognition using a spectrum. However, since the images are separated by direction and further subjected to blurring processing, the state in which each frame is normalized during the normalization is lost. This tendency is more pronounced in monosyllable recognition than in word recognition.

In other words, in word recognition, the vowel part (stable state) is long (many).
Therefore, the effect of blurring processing is less likely to appear, but in a single syllable, the preceding consonant part, that is, the part where the vector field changes,
This is thought to be because there are more words than words in terms of time. Therefore, good results can be obtained when the correlation coefficient used to calculate similarity is monosyllable recognition.

〔Effect of the invention〕

As described above, when using the method of the present invention, it is possible to achieve monosyllable recognition, which is extremely difficult to recognize, and even when the voice is uttered by an adult male, extremely high recognition can be achieved.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of a device for implementing the method of the present invention, Fig. 2 is a schematic diagram of a spatio-temporal pattern, Fig. 3 is a schematic diagram showing a normalized spatio-temporal pattern, and Fig. 4 is an explanatory diagram of vector direction quantization, and FIG. 5 is a schematic diagram of a two-dimensional pattern according to direction. 4... Feature extraction section 5... Blur processing section 6...
Standard pattern storage unit 7...Calculation unit 41...Normalization unit 42...Spectrum vector field extraction unit 43...
・Directional two-dimensional pattern creation unit patent Applicant: Director of the Agency of Industrial Science and Technology Satoshi Sugiura and one other agent and agent: Patent attorney Noboru Kono 1 Fig. Fig. Fig. 2 Fig.

Claims (1)

[Claims] 1. Obtain a spatio-temporal pattern of a scalar field defined by the temporal and spatial axes from an audio signal, and spatially differentiate the spatio-temporal pattern to have a magnitude and direction at each grid point in space. Convert it to a vector field pattern, quantize the direction parameter of the vector of the vector field pattern into N values (N: integer), separate this quantized value for each vector with the same value, and calculate the size of the vector. N direction-specific two-dimensional patterns are created with the value of each grid point as the value of each grid point, and a standard pattern created in advance using the direction-specific two-dimensional pattern and the direction obtained from the audio signal input as a recognition target are created. In a speech recognition method that identifies input speech by calculating the similarity with a feature pattern that is another two-dimensional pattern, the speech recognition is characterized in that the similarity is calculated for each of the directions of the standard pattern and the feature pattern. method. 2. The speech recognition method according to claim 1, wherein the similarity calculation for each direction is performed for each unit in the time axis direction of the two-dimensional pattern for each direction.