JPH0194397A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a speech recognition method that can obtain high recognition performance with a small number of learning patterns.

(Conventional technology) Inputting and outputting information through voice is highly natural for humans;
It is an excellent man-machine interface and has been studied in various ways. Most of the speech recognition devices currently in practical use are of the type that recognizes word speech, and the membrane is constructed as shown in FIG.

This device converts the input voice into an electrical signal and captures it, acoustically analyzes it in an acoustic analysis section 1 consisting of a band pass filter, etc., and detects the word speech section in a start/end detection section 2. do. Then, a pattern matching unit calculates the similarity and distance between the acoustic analysis data (feature information; speech pattern) of the word speech section of the input speech and each standard pattern of recognition target words registered in advance in the standard pattern dictionary 3. 4, and the result of the calculation is determined by the determining unit 5, to obtain, for example, the category name of the standard pattern with the highest similarity value as the recognition result for the input voice.

However, in speech recognition using the pattern matching method, a problem arises in that the input speech pattern and the standard pattern registered in advance are misaligned in the time axis direction (pattern deformation). Therefore, in the past, only linear expansion and contraction and dynamic programming (D
The problem with the shift in the time axis direction mentioned above is solved by nonlinear expansion and contraction represented by P).

On the other hand, apart from such a pattern matching method, an orthogonalized dictionary is created from learning patterns collected in advance.
A speech recognition method (subspace method) using this orthogonalized dictionary has been proposed. As shown in FIG. 3, an example of the configuration of this method is such that a sample point extraction unit 6 extracts a predetermined number of sample points obtained by equally dividing the speech section from a speech pattern that has been acoustically analyzed and detected a speech section. , (number of feature vectors ×
Find the sample pattern indicated by the number of sample points). A predetermined number of such sample patterns are collected for each category to be recognized and stored in the pattern storage section 7. Then, in the Gram-Schmidt (GS) orthogonalization unit 8, an orthogonalization dictionary 9 is created using the predetermined number (three or more) of sample patterns collected in the pattern storage unit 7 in the following procedure.

That is, in creating the orthogonalized dictionary 9, for each category, the m-th learning pattern of that category is set as ao, and when a learning pattern uttered three times is used, ■ the first learning data a1 is set as the first learning pattern. Let the axis dictionary b1 be b 1 ”” a t
(1) Register this in the orthogonalization dictionary 9.

(2) Using the Gram-Schmidt orthogonalization formula from the second learning data a2, perform the following calculation, and if l1b211 is larger than a certain value, register this in the orthogonalization dictionary 9 as the second axis dictionary b2. However, (.) indicates the inner product, T indicates the transposition, and 1111 indicates the norm.

■ Then, from the third learning data a3, perform the following calculation, and if 11b311 is larger than a certain value, use this as the third
It is registered in the orthogonalization dictionary 9 as the axis dictionary b3. However, if the second axis dictionary is not required, the above (2)
) calculates the formula.

The orthogonalized dictionary 9 is formed in advance by repeatedly performing the above processes (1) to (2) for each category.

The similarity calculation unit 10 calculates the category i between the orthogonalized dictionary 9 created as described above and the input speech pattern
, and the input speech pattern X is recognized according to this similarity value. Note that the above orthogonalized dictionary b of category i is normalized in advance to 1, r, and K1 is the number of dictionaries (number of axes) of category i.
It shows.

However, in a method using such GS orthogonalization, there is a problem that the amount of pattern variation carried by each of the above-mentioned orthogonal axes is not clear. For this reason, the sample pattern (a
r a + a l is i, l i, 2
1.3 There is a problem that it is not guaranteed that the original standard pattern of the category i is expressed well.

(Problems to be Solved by the Invention) In speech recognition using the conventional subspace method using GS orthogonalization, the orthogonalized dictionary itself contains, for example, the time axis direction of the collected learning patterns. There are problems caused by fluctuations in the direction of the frequency axis, and an issue remains as to whether the standard pattern is well represented. In addition, in order to solve such problems, there is a problem that it is necessary to collect a considerable amount of learning patterns.

The present invention was made in consideration of these circumstances, and its purpose is to create an orthogonalized dictionary that can adequately represent the standard pattern with a small number of learning patterns and can adequately cope with pattern variations. The object of the present invention is to provide a speech recognition method that can improve recognition performance.

[Structure of the Invention] (Means for Solving the Problems) The present invention provides a method for solving problems between an input speech pattern obtained by analyzing input speech and an orthogonalized dictionary created based on learning patterns collected in advance. In the speech recognition method that recognizes the input speech by calculating the similarity with Find the average pattern of , smooth this average pattern in the time axis direction and the frequency axis direction to find the first axis dictionary, further differentiate the above average pattern in the time axis direction to find the second axis dictionary, and , create the orthogonalized dictionary by differentiating the average pattern in the frequency axis direction to obtain a third axis dictionary, and further create an additional dictionary orthogonal to the orthogonalized dictionary by, for example, Gram-Schmidt orthogonalization. The additional dictionary is created and added to the orthogonalized dictionary.

(Operation) The average pattern of the learning patterns collected using three or more types of filters is obtained, and this average pattern is smoothed in the time axis direction and the frequency axis direction to obtain the first axis dictionary. It is possible to effectively absorb axial fluctuations and frequency axial fluctuations.

Furthermore, since the average pattern is differentiated in the time axis direction to obtain a dictionary on the second axis, it is possible to effectively absorb the positional shift of the audio pattern in the time axis direction, and the above average pattern is differentiated in the frequency axis direction. Since the dictionary for the third axis is obtained by using the above method, it is possible to effectively absorb the positional deviation of the voice pattern in the direction of the frequency axis.

In this way, an orthogonalized dictionary is created that absorbs pattern fluctuations in the time axis direction and frequency axis direction, so each dictionary pattern in the orthogonalized dictionary can be made to be able to cope with positional shifts due to the fluctuations. , greatly contributes to improving recognition performance. Furthermore, dictionaries for the second and third axes are obtained based on a dictionary pattern (first axis) generated from an average pattern that absorbs pattern fluctuations in the time axis direction and frequency axis direction, and the orthogonalized dictionary is generated. Therefore, unlike in the past, the amount of pattern variation carried by each orthogonal axis of the orthogonal dictionary itself does not become unclear, and it is possible to effectively create a high-performance orthogonal dictionary by effectively using a small number of learning patterns. becomes possible.

Furthermore, since an additional dictionary orthogonal to the orthogonalized dictionary is created and added to the orthogonalized dictionary, this additional dictionary can effectively absorb pattern fluctuations other than those in the time axis direction and frequency axis direction. This makes it possible to perform recognition processing using the same method, which greatly contributes to improving recognition performance.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a speech recognition device constructed by applying an embodiment method according to the present invention, and the same parts as those of the conventional device shown in FIG. 3 are denoted by the same reference numerals. .

The feature of this embodiment device is that, as a means for creating an orthogonalization dictionary 9 using the learning patterns accumulated in the pattern storage section 7, at least smoothing processing and differential processing are performed instead of the conventional GS orthogonalization section 8. The present invention uses an orthogonal time/frequency filter section 11 that is composed of three or more types of filters, for example, orthogonal time/frequency filters. Furthermore, for example, using the GS orthogonalization section 8, a dictionary orthogonal to the orthogonalized dictionary created in the orthogonalization time/frequency filter section 11 is created as an additional dictionary, and this additional dictionary is used for the orthogonalization. The feature is that it is added to the dictionary 9.

Here, the learning pattern collected in the pattern storage section 7 is composed of a feature vector of 16 points indicated by j (-1, 2, ~1B), which has not been subjected to acoustic analysis, and its speech interval is defined as k (- 0,4,2,~17) as 17
The explanation will be given assuming that it is given as a data series collected over 18 equally divided sampling points.

Therefore, the orthogonalization time/frequency filter unit 11 converts the orthogonalization dictionary 9 into gray, for example, when the m-th learning pattern collected in groups of three is aa+(j, k) for the category i. Creating.

■ First, from the learning pattern am(j,k) of category i, its average pattern AU,k) is N-0,1,2,~15. k-0,1,2,~1f].

■ After that, the average pattern A obtained as described above
Using (Lk) bl(j,k) +A +A ” A(j-1, to-1) U-1,k)
(j-1, +1) + 2*A +A ” A (j, -1) (j, k) (
+1 to j) +A +A +A (-1 to j+1.) (j+1.k)
(+1 to j+1.) [j-1, 2, ~ 14.
k-1, 2, ~ 1θko・
The dictionary b1N,k) of the first axis is obtained by the calculation (6) and is registered in the orthogonalization dictionary 9. This dictionary bl(j
, k) are obtained by smoothing the average pattern A(j, k) in the time axis direction and the frequency axis direction, respectively, and are registered as first axis dictionary data serving as a reference for the orthogonalized dictionary 9.

■ After that, using the average pattern A(j, k),
−-A +A b2(j, k) (j-1, to-1) (j
−1, +1)” = A(j, k() (j, +1)” = A(j, k() (j, +
1) 'ten A ” ” U+1. −1) (+1 to j+1.)
'+ A [j=1.2. ~14. k-1, 2, ~1B]
...(7) With the calculation, the dictionary b of the second axis
2(j,k) is obtained, normalized, and then registered in the orthogonalization dictionary 9.

This dictionary b2(j,k) is the average pattern A(j,k
) is obtained by differentiating it along the time axis.

Note that the second axis dictionary b2(j,
k) is not completely orthogonal to the first axis dictionary b1(j, k), so B2(j,
k)″b2(j,k)′ ” 2(j,k) 1(j,k))51(j,
After performing re-orthogonalization processing such as k) and b and normalizing this re-orthogonalized dictionary data B2 (j, k), the orthogonalized dictionary is used as a new second axis dictionary b2 (j, k). 9 may be registered. However, even without such re-orthogonalization, the second axis dictionary b2(j,
It is possible to obtain sufficient recognition performance with k).

■ Also, using the average pattern AU, k),
-A b3 (j, k) (j-1, to -1) (j
-1, k) + A -A (+1 to j-1) (-1 to j+1.) +
A +A(j+1.k) (+1 to j+1.) [
j-1, 2, ~14. k-1, 2, ~16]
... (8) Dictionary b3 of the third axis with the calculation
After finding (j, k) and normalizing it, it is registered in the orthogonalization dictionary 9.

This dictionary b3(j,k) is the average pattern A(j,k
) is obtained by differentiating it in the frequency axis direction.

The orthogonalized dictionary 9 is created by repeatedly performing the above-mentioned processes (1) to (2) for each category.

In the above explanation, an example was shown in which up to three axes are obtained as the orthogonal dictionary 9, but dictionaries for four axes or later may be created by further performing second-order differentiation or the like. In this case, instead of the above-mentioned 18 points as a learning pattern, for example, a pattern obtained by extracting 20 or more sample points may be used.

On the other hand, the GS orthogonalization unit 8 creates an additional dictionary orthogonal to the orthogonal dictionary from the learning patterns collected in the pattern storage unit 7 in the following manner.

That is, the GS orthogonalization unit 8 performs Gram-Schmidt orthogonalization of the learning pattern an(j, k) collected in the pattern storage unit 7, where P is the number of axes of the orthogonalization dictionary that has already been obtained. Calculating an expression.

If 11b 11 is larger than a predetermined value, pm is registered in the orthogonal dictionary 9 as an additional dictionary. The creation of this additional dictionary is performed sequentially for the learning patterns am(j, k) stored in the pattern storage section 7.

In this way, an orthogonalized dictionary set consisting of an orthogonalized dictionary created by smoothing and differentiation using an orthogonalized time/frequency filter, and an additional dictionary obtained by Gram-Schmidt orthogonalization based on this orthogonalized dictionary is created. According to this device, which recognizes and processes an input speech pattern, the orthogonalization dictionary 9 absorbs variations in the speech pattern in the time axis direction and frequency axis direction, and further absorbs other pattern variations. Since the input speech pattern has also been absorbed, speech recognition can be performed without being affected by fluctuations in the time axis direction and frequency axis direction of the input speech pattern, and the recognition performance can be improved. In addition, the orthogonalized dictionary 9 is created using the orthogonalized time/frequency filter, and it becomes possible to efficiently construct a high-performance orthogonalized dictionary with a small number of learning patterns, which has a great practical effect.

According to this method, speech recognition is performed by creating an orthogonalized dictionary using a differential filter that compensates for positional deviations in the time axis direction and frequency axis direction and an orthogonalization filter that absorbs fluctuations in two-dimensional patterns. For example, it can be seen that high recognition performance can be obtained with a small number of learning patterns. Furthermore, the additional dictionary allows pattern variations other than those described above to be effectively absorbed for speech recognition. Therefore, it can be said that this method has a great effect on improving speech recognition performance.

Note that the present invention is not limited to the embodiments described above. Although an example of creating a three-axis orthogonal dictionary has been described here, it is also possible to create an orthogonal dictionary with a larger number of axes. In this case, several variations can be considered for the coefficients of the orthogonalized time/frequency filter, but
In short, the learning pattern can be smoothed, first differentiated, second differentiated, etc. in the time axis direction and the frequency axis direction, and can be implemented with various modifications. Furthermore, the number of additional dictionaries (number of axes) added to the orthogonalized dictionary is not particularly limited. Furthermore, the number of dimensions of the learning patterns used to create the dictionary is not particularly limited. Furthermore, it is also possible to create an additional dictionary using an orthogonalization method other than Gram-Schmidt orthogonalization. In addition, the present invention can be modified and implemented without departing from the gist thereof.

[Effects of the Invention] As explained above, according to the present invention, an orthogonalized dictionary that absorbs pattern fluctuations in the time axis direction and pattern fluctuations in the frequency axis direction is created using three or more types of filters, and the orthogonalized dictionary is Since an additional dictionary that is orthogonal to the dictionary is created, it is possible to obtain a dictionary that effectively expresses variations in the patterns with a small number of learning patterns, and has great practical effects such as improving recognition performance.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram of a speech recognition device constructed by applying an embodiment of the present invention, and FIGS. 2 and 3 are diagrams each showing a schematic configuration of a conventional speech recognition device. ■...Acoustic analysis section, 2...Start/end detection section, 5.
... Judgment unit, 6... Sample point extraction unit, 7... Pattern storage unit, GS orthogonalization unit, 9... Orthogonalization dictionary, 1
0... Similarity calculation unit, 11... Orthogonal time/frequency filter. Applicant's agent Patent attorney Takehiko Suzue

Claims (3)

[Claims] (1) Speech recognition that recognizes the input speech by calculating the similarity between the input speech pattern obtained by analyzing the input speech and an orthogonal dictionary created based on the learning patterns collected in advance In the method, means for creating the orthogonalized dictionary using three or more types of filters that perform at least smoothing processing and differentiation processing on learning patterns collected in advance, and creating an additional dictionary that is orthogonal to the orthogonalized dictionary. A voice recognition method characterized by comprising means for. (2) The filter includes means for obtaining an average pattern of the collected learning patterns, smoothing this average pattern in the time axis direction and frequency axis direction to obtain a first axis dictionary, and smoothing the average pattern in the time axis direction. Speech recognition according to claim 1, comprising means for differentiating the average pattern to obtain a second axis dictionary, and means for differentiating the average pattern in the frequency axis direction to obtain a third axis dictionary. method. (3) The speech recognition system according to claim 1, wherein the means for creating the additional dictionary creates an additional dictionary that is orthogonal to the orthogonalized area by Gram-Schmidt orthogonalization.