JPH0449719B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device capable of recognizing both monosyllabic speech and word speech.

Structure of the conventional example and its problems A speech input word processor that uses a monosyllabic speech recognition device that recognizes monosyllables generated by dividing them into words instead of the kana keyboard of a word processor with a kana-kanji conversion function has already been commercialized. ing.

In this case, voice recognition is for so-called specific speakers, in which the user must register his/her own voice as a standard pattern in advance. At this time,
In order to use function functions such as "line break", "deletion", "conversion", and other frequently occurring words as word sounds, it is necessary to use feature vectors not only as standard patterns for word clauses but also as standard patterns for words. The series must be registered. For this reason, as the number of words increases, the storage capacity increases, and the amount of calculation also increases significantly, as it is necessary to calculate the distance between the feature vector of the input pattern and the feature vector series of monosyllables and words. There was a problem.

OBJECTS OF THE INVENTION An object of the present invention is to solve the above problems and provide a speech recognition device that can recognize monosyllables and words while suppressing increases in storage capacity and calculation amount.

Structure of the Invention The speech recognition device of the present invention registers monosyllable standard patterns in advance, recognizes monosyllables uttered in sections, and uses the word standard pattern as a combination pattern of the monosyllable standard patterns, for each monosyllable. The system is configured to perform word recognition using the distance between vectors determined by the method, so that both monosyllable recognition and word recognition can be performed.

DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of the present invention will be described, but before that, a word speech recognition device using pattern matching will be described. The general configuration of this device is as follows.

Input audio signal, filter bank, frequency analysis
Feature extraction means for converting into a series of feature vectors by LPC analysis etc., and standard pattern storage means for registering the series of feature vectors uttered in advance and extracted by the feature extraction means as standard patterns for all recognized words. and pattern comparison means for calculating the similarity or distance as a series of feature vectors between the input pattern generated for recognition and extracted by the feature extraction means and all of the standard patterns stored in the standard pattern storage means. and a determining means for determining and outputting a word corresponding to the standard pattern having the highest degree of similarity (smallest distance) as a recognition result as a result of pattern comparison.

At this time, even if the same speaker utters the same word, the duration of the utterance differs each time, so when comparing the standard pattern and the input pattern using the pattern comparison means, the time axis of both It is necessary to expand and contract the pattern lengths of the two to make them the same and compare them. At this time, since the change in utterance time length does not occur uniformly in each part of the uttered word, it is necessary to expand and contract each part non-uniformly.

Figure 1 represents this graphically. In Figure 1a, the horizontal axis is the i coordinate corresponding to the input pattern T = a ₁ a ₂ ... a _I (a _i is the feature vector of the i-th frame of the input pattern), and the vertical axis is the standard pattern R ⁿ = b ⁿ ₁ b ⁿ ₂
... represents the j coordinate corresponding to b ⁿ _J n (b ⁿ _j is the feature vector of the j-th frame of the standard pattern R ⁿ ). Matching the input pattern T and the standard pattern R ⁿ by nonlinearly expanding and contracting the time axis means that on this lattice graph, the path 1 that tightens the correspondence of each feature vector of both patterns is defined as a series of both patterns. Find out based on the evaluation criterion that the distance is the minimum,
The distance at that time is defined as the distance between both patterns. A method using dynamic programming is well known as a method for efficiently performing this calculation, and is called DP matching.

When determining this route, limiting conditions are set in consideration of the nature of the voice. FIG. 1b is an example of a route selection condition called slope restriction. That is, in this example, the path leading to point (i, j) is point (i-2, i-
1) through point (i-1,j) is point (i
-1,j-1) or point (i-1,j-
This means that only one path can be taken from 1) to the point (i, j-1), and if we provide the condition that the starting and ending ends of the input pattern and the standard pattern must correspond, then the above The matching path is limited to the shaded area in FIG. 1a. This restriction applies no matter how the time axis expands or contracts.
This is to prevent extreme correspondences from occurring due to the fact that the same word cannot be expanded or contracted so drastically.

The distance between both series is defined as a weighted average along the path of the intervector distance d ⁿ (i, j) between the input vector a _i and the standard pattern vector b ⁿ _j .

At this time, the DP matching method can be used if the sum of the weights along the path is made constant regardless of how the path is selected.

FIG. 2 illustrates the matching of the word standard pattern constructed by combining end syllable speech standard patterns and the input pattern.
In the same figure, R ^q (1), R ^q (2), and R ^q (3) are monosyllables q(1),
It means the standard pattern of q(2), q(3), and this example shows the case of matching the input pattern with the standard pattern of words consisting of monosyllables q(1), q(2), q(3). ing. According to the above description, the matching path is, for example, 2.

Examples of the present invention using the pattern matching method described above will be described below.

FIG. 3 is a block diagram showing one embodiment of the present invention. In the figure, 3 is an audio signal input terminal;
is a feature extractor that converts an input audio signal into a series of feature vectors, which is composed of a filter bank or the like. Reference numeral 5 denotes a monosyllabic standard pattern storage unit in which patterns converted into a series of feature vectors for each monosyllable are stored. Reference numeral 6 denotes an inter-vector distance calculation unit which calculates the distance d ⁿ (i, j) between the vector b ⁿ _j forming the standard pattern R ⁿ in the monosyllabic standard pattern storage unit 5 and the vector a _i forming the input pattern. do.
a _i = (a _i1 , a _i2 , ..., a _il ), b ⁿ _j = (b ⁿ _j1 , b ⁿ _j2 , ...,
b ⁿ _jl )
Then, d ⁿ (i, j) is most simply given by d ⁿ (i, j) = _l 〓 ^k=1 | a _ik − b ⁿ _jk |.

Reference numeral 7 denotes an inter-vector distance storage unit which stores the results calculated by the inter-vector distance calculation unit 6.

Reference numeral 8 denotes a monosyllable cumulative distance calculation unit, which calculates the cumulative distance (weighted sum along the matching path of d ⁿ (i, j)) from the first frame to the current frame for each monosyllable. If Figure 1b is adopted as the constraint condition for the matching route, and the weighting coefficient along each route is the numerical value attached to the route in the same figure, then the coordinates (i,
Cumulative distance for standard pattern R ⁿ in j)
D ⁿ (i, j) is given as follows.

D ⁿ (i, j) = minD ⁿ (i-2, j-1) + d (i-
1,j)+d(i,j) D(i-1,j)+d(i,j) D(i-1,j)+d(i,j) D ⁿ (i-1,j-2)+0 .5d(i-1,j)+0.5d(i
, j)...(1) 9 is a monosyllable determination unit, which recognizes the monosyllable corresponding to R ⁿ when n where D ⁿ (I, J ⁿ ) is minimum is set to n^. .

Reference numeral 10 is a word dictionary in which recognized words are prepared by inputting them using a keyboard or the like as a string of monosyllabic symbols representing the words. Reference numeral 11 denotes a word cumulative distance calculation unit, which stores words to be matched in the inter-vector distance storage unit 7 in the order of monosyllables specified in the word dictionary 10. The already calculated distance between vectors is read out, and the cumulative distance to the point (i, j) as a word is calculated. That is, for example, in FIG. 2, in the i-th frame, the monosyllabic standard pattern R ⁿ = b ⁿ ₁ b ⁿ 2 ... b ⁿ _Jo for n = 1, _2, ...N (N is the number of monosyllables)
each vector b ⁿ _j and input pattern T=a ₁ a ₂
...The inter-vector distance d ⁿ (i, j) between vector a _i of the i-th frame of a _I has already been calculated during monosyllable recognition, so R ^q(1) , R ^q(2) , R combination pattern of ^q(3)
R ^q(1) R ^q(2) R ^q(3) =b ^q(1) ₁ b ^q(1) ₂ …b ^q(1) _Jq(1) b ^q(2)
₁ b ^q(2) ₂ …b ^q(2) _Jq(2)
b ^q(3) b ^q(3) ...b ^q(3) There is no need to newly calculate the vector distance between _Jq(3) and a _i . After the input is completed, the word determination section 12 determines the one that gives the minimum value among the final cumulative distances for each word obtained by the word cumulative distance calculation section 11 as the word recognition result. Reference numeral 13 denotes a monosyllable/word identification unit, which identifies whether the input voice is a monosyllable or a word. This can be specified from the keyboard, but it can also be determined automatically based on the length of the voice section. In other words, since words take longer to occur than monosyllables, it is possible to determine whether a speech interval obtained from the power of the input signal using a normal method exceeds a predetermined threshold or not to determine whether it is a word or a monosyllable. . Reference numeral 14 denotes a recognition result switching unit that switches the output of the monosyllable determining unit 9 or the word determining unit 12 according to the result of the monosyllable/word identifying unit 13. 15 is an output terminal for the recognition result.

As in this embodiment, if the calculation of the inter-vector distance and cumulative distance is completed for every single syllable and every word in each input frame, the recognition result can be obtained at the same time as the input ends. At this time, if the constraint condition for route selection is shown in FIG. It is sufficient to store only the i-th frame and the i-1th frame. Further, regarding the cumulative distance, it is sufficient to store only those for the i-1th frame and the i-2th frame, as well as those for single syllables and words. Further, regarding cumulative distances, it is sufficient to store only those for the i-1th frame and the i-2th frame, both for single syllables and for words. Note that storage units for these cumulative distances are included in the monosyllable cumulative distance calculation unit 8 and the word cumulative distance calculation unit 11, and are not shown.

In this embodiment, the determining units are divided into monosyllables and words, such as the monosyllable determining unit 9 and the word determining unit 12, and the monosyllable/word identifying unit 9 determines whether the input speech is a monosyllable or a word. Although the recognition result switching unit 14 outputs the determination result of either of the determination units, in another embodiment, only one determination unit is provided and the monosyllable cumulative distance calculation unit It is also possible to output a single syllable or a word corresponding to the standard pattern with the smallest distance as a recognition result without distinguishing between the output of the word cumulative distance calculating section 11 and the output of the cumulative word distance calculating section 11. In that case, the monosyllable/word identification unit 13 and the recognition result switching unit 14 are unnecessary.

Note that the functions of each component of the embodiment described above can also be realized by software means.

Effects of the Invention The speech recognition device of the present invention registers monosyllabic standard patterns in advance, recognizes monosyllables generated in sections, and uses word standard patterns as a combination pattern of the monosyllabic standard patterns, for each monosyllable. Since the word cumulative distance is calculated using the distance between the vectors found in There is no need to register word sounds as standard patterns in addition to syllable sounds, just register monosyllabic sounds to create words that correspond to faction keys,
Frequently appearing words can also be input using word recognition, which is highly valuable.

[Brief explanation of the drawing]

Figures 1a and b are diagrams explaining the principle of DP matching, Figure 2 is a diagram explaining the principle of the present invention, and Figure 3 is a diagram explaining the principle of DP matching.
The figure is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention. 4... Feature extraction unit, 5... Monosyllabic standard pattern storage unit, 6... Inter-vector distance calculation unit, 7... Inter-vector distance storage unit, 8... Monosyllabic cumulative distance calculation unit, 9... Monosyllabic Judgment unit, 10...Word dictionary, 1
1...Word cumulative distance calculation unit, 12...Word determination unit.

Claims (1)

[Claims] 1 Input audio signal as a sequence of feature vectors a ₁ a ₂ …a _i
...a _I , a feature extraction means for converting the monosyllabic speech signal of the n-th monosyllabic into an input pattern T consisting of a feature vector sequence b ⁿ ₁ b ⁿ ₂ ...b ⁿ _J n (where n=1, 2, ...,
A standard pattern storage means for storing the standard pattern R ⁿ converted into N), and a feature vector ^{b n} _j (j = 1, 2, ..., J ⁿ ) and the feature vector a _i of the i-th frame of the input pattern T, inter-vector distance d ⁿ (i, j)
and an inter-vector distance calculation means for calculating the input pattern T from the inter-vector distance d ⁿ (i, j).
and monosyllabic cumulative distance calculation means for calculating the distance between the series of the standard pattern R ⁿ and the word standard pattern expressed as a combination of the monosyllabic standard pattern and the series of the input pattern. From the results of the cumulative distance calculation means and the monosyllable cumulative distance calculation means or the word cumulative distance calculation means,
A speech recognition device characterized by comprising: determining means for determining a single syllable or word standard pattern that is closest in distance to the input pattern T.