JPH0225899A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To easily prepare an orthogonalized dictionary to deal with a speaker by making the dictionary of an orthogonal axis obtained first from the learning patterns of plural speakers into a basis and successively dictionary-registering an orthogonal vector according to the learning pattern obtained from the individual speaker. CONSTITUTION:First, according to the learning patterns to be respectively obtained from plural speakers, the orthogonal axis made into the basis is decided, a first dictionary is prepared, and it is dictionary-registered to an orthogonalized dictionary 9. Next, as to a dictionary preparation according to the learning pattern obtained from the individual speaker, the dictionary of the axis to be respectively orthogonal to each axis of the dictionary to be already registered is obtained, whether to register the dictionary or not is successively decided, it is additionally registered to the orthogonalized dictionary 9, and thereby, the orthogonalized dictionary 9 is constructed. Thus, the orthogonalized dictionary with a high recognizing performance can be efficiently prepared.

Description

[Detailed Description of the Invention] [Object of the Invention] (Field of Industrial Application) The present invention can be commonly used by members of a specific group,
The present invention relates to a speech recognition device that can obtain high recognition performance with a small number of learning patterns from group members.

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

This device converts the voice input into an electrical signal and captures it.
The sound analysis section 2 performs acoustic analysis, and the start/end detection section 2 detects the speech section of the word. 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 stretching and dynamic programming (DP) were used.
) The above-mentioned problem with the deviation in the time axis direction is solved by non-linear expansion and contraction as typified by.

On the other hand, apart from such a pattern matching method, there is a method (subspace method) in which an orthogonalized dictionary is created from learning patterns collected in advance and speech recognition is performed using this orthogonalized dictionary.
has been proposed. As shown in FIG. 5, 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. Then, a sample pattern represented by [number of feature vectors x number of sample points] is obtained. A predetermined number of such sample patterns are collected for each category to be recognized and stored in the pattern storage section 7. And Gram
In the Schmidts I-(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, to create the above-mentioned orthogonalized dictionary 9, for each category, the m-th learning pattern of that category is set as al, l, and when using the learning pattern that has been uttered three times, ■ the first learning data a1 is Let the dictionary b1 be the first axis, and 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 1lb211 is larger than a certain value, register this in the orthogonalization dictionary 9 as the second axis dictionary b2. However, (.) indicates the inner product, and 1111 indicates the norm.

■ Then, from the third learning data a3, perform the following calculation, and if 1lb311 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 IO is an orthogonalization unit created as described above.
9 and human voice pattern X, category i
This is to calculate the degree of similarity between the orthogonalized dictionary b and the orthogonalized dictionary b. The input speech pattern X is recognized according to the similarity value determined for each of these categories i. In addition, the orthogonalized dictionary b of the above category i
is pre-normalized 1, ``, and K1 indicates the number of dictionaries (number of axes) of category i.

By using such GS orthogonalization, the recognition performance is greatly improved. Furthermore, attempts have been made to create orthogonalized dictionaries that absorb fluctuations in the time and frequency directions using differential filters, and to further improve their recognition performance.

However, in this type of device, a standard speech dictionary is created exclusively for a specific speaker.

For this reason, when another speaker attempts to use the speech recognition device 17, it is necessary to change the speech dictionary each time. Therefore, it has been considered to collect many learning patterns from many speakers and create an orthogonal dictionary.
Problems such as creating a dictionary became unnecessarily complicated and making it difficult to obtain a dictionary with high recognition performance occurred.

(Problems to be Solved by the Invention) In speech recognition using the subspace method using conventional orthogonalized dictionaries, it is difficult to determine how to obtain high-performance orthogonal dictionaries from learning patterns collected from multiple speakers. There remains work to be done on how to efficiently create a dictionary. There is also a problem in how to efficiently collect the learning patterns of multiple speakers necessary for creating an orthogonal dictionary and create an orthogonal dictionary.

The present invention has been made in consideration of these circumstances, and its purpose is to efficiently create an orthogonalized dictionary with high recognition performance using a small number of learning patterns collected from a plurality of speakers. An object of the present invention is to provide a speech recognition device with high recognition performance that can be commonly used by a plurality of users.

[Structure of the Invention] (Means for Solving the Problems) The present invention is constructed based on input speech patterns obtained by analyzing input speech and learning patterns collected in advance. In the speech recognition device that recognizes the input speech by calculating the similarity between the input speech and the orthogonalized dictionary, the orthogonalized dictionary determines a basic orthogonal axis from the learning patterns of multiple speakers and serves as a reference dictionary. After creating a dictionary and registering it in a dictionary, when creating a dictionary from the learning patterns of individual speakers, determine a new axis that is orthogonal to the width of the already registered dictionary, and add the dictionary of this new axis to the dictionary. The feature is that whether or not to register is determined based on, for example, the value of the norm of the axis, and the orthogonalized dictionary is sequentially constructed.

(Operation) According to the present invention, orthogonal axes, which are the basis of an orthogonalized dictionary, are determined from learning patterns obtained from a plurality of speakers, a dictionary is created, and after the dictionary is registered, individual When creating a dictionary based on learning patterns from speakers, an axis that is perpendicular to the axis of the dictionary that has already been created and registered in the dictionary is found, and a dictionary about this new axis is created from the learning patterns of the individual speakers. It will be done. Then, by determining the value of the norm, it is checked whether or not it should be additionally registered in the dictionary, and the orthogonalized dictionary is additionally registered in the dictionary only when it is useful as a dictionary.

As a result, it becomes possible to construct an orthogonalized dictionary that efficiently expresses the pattern variation elements from the learning patterns of a plurality of speakers, and it becomes possible to obtain an orthogonalized dictionary with high recognition performance. Moreover, since it is possible to efficiently incorporate the variation patterns of individual speakers into a set of orthogonal vectors and express the dictionary in relation to the basic dictionary of orthogonal axes, the amount of calculation can be reduced and the dictionary can be easily and efficiently used. It becomes possible to create.

(Example) Hereinafter, one 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 according to an embodiment of the present invention, and the same parts as those of the conventional device shown in FIG. 5 are denoted by the same reference numerals.

This embodiment device is characterized by an orthogonal vector calculation unit 8a, an orthogonal vector registration determination unit 8b, and a residual An orthogonalized dictionary creation unit 8 is provided which includes a norm memory 8c, and the processing concept in this orthogonalized dictionary creation unit 8 is schematically shown in FIG.
First, the basic orthogonal axes are determined according to the learning patterns obtained from a plurality of speakers, the first dictionary is created, and the dictionary is registered in the orthogonalization unit w9, and then the dictionary is created according to the learning patterns obtained from each individual speaker. To create a dictionary, find a dictionary with axes orthogonal to the basic axis (orthogonal to each axis of the already registered dictionaries), and create an orthogonalized dictionary by sequentially determining whether to register this dictionary or not. The feature is that an orthogonalized dictionary 9 with high recognition performance is constructed by adding additional registrations to the dictionary 9.

The creation of the orthogonalized dictionary in the orthogonalized dictionary creation section 8 will be explained in more detail according to the processing procedure shown in FIG.

In this case, the learning pattern collected in the pattern storage unit 7 is composed of six acoustically analyzed feature vectors indicated by, for example, j (-1, 2, ~8), and the speech interval is defined as k (- 0,1,2,~l l) and 5 and 11
The explanation will be given assuming that it is given as a data series collected over 12 equally divided sampling points.

The orthogonalized dictionary creation unit 8 first collects three learning patterns from each of a plurality of (L) speakers for category i to be registered in the dictionary (step a). After that, the m-th learning pattern ([T1-1゜2.3.~M; M-3
x L) is defined as am(j, k), then the basic orthogonalization dictionary 9 is created as follows.

■ First, from the learning pattern am(j-k) of category i, find its average pattern A as (j, k) [j-1, 2, ~1B, -0, 1, 2, ~17] ( Step b).

■ After that, using the average pattern A(j, k) obtained as described above, -A +2*A +Abl(j, k
) (j, ni-1) (j, k)
(j,+1) [j=1,2. ~16. k-L2
．． ~16th... Find the dictionary bl(j, k) of the first axis by the calculation (6) (step c) and register it in the orthogonalization dictionary 9 (step d)
. This dictionary b is obtained by smoothing the average pattern 10, k) A(j, k) in the time axis direction, and becomes dictionary data on the first axis that serves as a reference for the orthogonalized dictionary 9.

■ After I2, using the average pattern A (j=k), calculate -A +A b2(j, k) (j, -1) (j, +1) [j=1,2. ~1B, k-1, 2, ~16]
. . . In the calculation (7), a dictionary b2(j, k) of the second axis of C is obtained (step e), and after being normalized, it is registered in the orthogonalization dictionary 9 (step f). This second axis dictionary b2(j,k) is the average pattern A(
C is obtained by differentiating j, k) in the time axis direction.

Incidentally, the second axis dictionary b2(j,
k) is not completely orthogonal to the first axis dictionary b1(j,k), so "2U,k)"b2(j,k) (b2(j,k) 1( , i, k)) bl(j, k
) Conflict b After performing re-orthogonalization processing and normalizing this re-orthogonalized dictionary data B2 (j, k), a new second axis dictionary b
2(j,k) may be registered in the orthogonalization dictionary 9 as 2(j,k).

Also, here we have shown an example of creating up to the second axis, but there are also two
Of course, it is also possible to create dictionaries for the third and subsequent axes as orthogonalized dictionaries for the basic axes by performing second-order differentiation or the like.

■ Thereafter, based on the orthogonalized dictionary obtained as described above, the orthogonal vector calculation section 8a sequentially extracts the individual learning patterns of a plurality of speakers stored in the pattern storage section 7 (step g). ), and according to the learning pattern, an additional dictionary that is orthogonal to the above orthogonalized dictionary is created as follows.
Create with 7.

That is, the creation of this additional dictionary is based on the learning pattern al(j, k) collected in the pattern storage section 7, where P is the number of axes of the orthogonalized dictionary already obtained [n = 1.2 ．． ~p, m-1,2,~M
This is performed by calculating the Gram-Schmidt orthogonalization formula l (step h). Then, the orthogonal vector (additional dictionary) b expressing this newly found characteristic variation of each individual speaker
is inputted to the orthogonal vector registration determination unit P 8b, and it is determined whether the norm llbII is greater than a predetermined value (step i). If the norm value is larger than a predetermined value, this is used as an additional dictionary to perform pattern normalization processing, and then the orthogonalized dictionary 9
(Step j). In this case, the above norm llb
The value of 11 is registered in the residual error pm - 8c (step k).

The orthogonalized dictionary 9 for category i is created by repeatedly performing the above processes 1 to 2 for each individual learning pattern obtained from a plurality of speakers.

Note that when registering a dictionary of newly obtained axes in the orthogonalization unit lf9, the number of axes that is predetermined as the orthogonalization dictionary 9 may be exceeded. In such a case, if dictionary registration of a new axis is stopped, there is a risk that the recognition performance for the speaker who obtained the dictionary will deteriorate. Therefore, in such a case, the category i is stored from the residual norm memory 8c.
The residual norm llb II on each axis for P++11 is read out and compared with the value of the residual norm on the new axis. If there is a dictionary whose residual norm value is small among the already registered dictionaries, the dictionary (orthogonal vector) corresponding to that residual norm is deleted from the orthogonalized dictionary 9 and replaced with the above-mentioned new dictionary. The obtained dictionary (orthogonal vector) is registered in the dictionary.

In this case, it goes without saying that the value of the corresponding norm in the residual norm memory 8c is also rewritten.

Based on the dictionary of orthogonal axes that is first obtained from the learning patterns of multiple speakers as described below, orthogonal vectors according to the learning patterns obtained from each individual speaker are sequentially registered in the dictionary to create an orthogonalized dictionary. 9 will be built. As a result,
If the number of speakers is within a certain range, it is possible to obtain an orthogonalized dictionary 9 with high recognition performance for the input speech patterns of all registered speakers, and it is possible to improve the recognition performance. Become.

As mentioned above, new axis dictionaries are created one after another through simple arithmetic processing, so the processing load is very light, and the orthogonalized dictionary 9 that can be adapted to multiple speakers can be created efficiently. Effects such as making it possible to do the following can be achieved.

The following table shows the voice data of 30 words consisting of numeric sounds and personal names collected from 5 men, 2 women, and 3 women over 13 times each.
This shows an experimental example in which 3 of the tests were used for learning and the remaining 10 tests were used for recognition performance evaluation.

Note that speaker A in this table is a speaker with relatively poor performance, and speaker B is a speaker with good performance. These results also show the recognition performance when using the orthogonalized dictionary SENODO at the time when all 10 speakers have completed dictionary registration. For reference, the recognition performance when the above speakers A and B create an orthogonalized dictionary (4 axes) with so-called specific speakers is as follows.
They were 92.5% and 98.3%, respectively.

As shown in this experimental data, according to this method, 10
By creating the orthogonalized dictionary 9 as described above for registered speakers of approximately 500,000 registered speakers, it is possible to perform speech recognition stably and with relatively high performance for all registered speakers regardless of the registration order. became clear.

Note that the present invention is not limited to the 17 embodiments described above. First, an example will be described in which a two-axis orthogonalized dictionary is created from a plurality of registered speakers, but it is also possible to create a basic orthogonalized dictionary with an even larger number of axes. In this case, several variations can be considered for the coefficients of the orthogonalization filter, but the key is to smooth the learning pattern, differentiate it to the first order, and differentiate it to the second order.

. . . and can be implemented with various modifications. Furthermore, the number of dimensions of the learning pattern is not particularly limited either. Furthermore, the number of axes of a newly created dictionary may be determined according to the number of learning patterns, and it is also possible to create a dictionary using an orthogonalization method other than Gram-Schmidt. 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, learning patterns collected from a plurality of speakers are used to easily generate n orthogonalized dictionaries compatible with these speakers with good performance. Therefore, it is possible to obtain a dictionary that effectively expresses pattern variations with a small number of learning patterns, and this 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 according to an embodiment of the present invention, FIG. 2 is a diagram schematically showing the concept of orthogonal dictionary creation in the embodiment device, and FIG. 3 is a diagram showing the concept of orthogonal dictionary creation in the embodiment device. FIGS. 4 and 5 are diagrams showing an example of a processing procedure, and each is a diagram 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 . ... Orthogonal vector calculation unit, 8b... Orthogonal vector registration determination unit, 8c... Residual norm measurement.

Claims (1)

[Claims] The similarity is calculated between an input speech pattern obtained by analyzing input speech and an orthogonal dictionary created based on learning patterns of multiple speakers collected in advance. In the speech recognition device that recognizes the input speech, after determining the orthogonal axes that are the basis of the orthogonalized dictionary from the learning patterns of multiple speakers, dictionary creation from the learning patterns of individual speakers is performed using the already registered A speech recognition device characterized in that the orthogonalized dictionary is constructed by determining a new axis orthogonal to an axis of the existing dictionary, and determining whether or not to register the dictionary of this new axis.