JPH0120439B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention relates to a recognition method for input speech in which a plurality of words or syllables are consecutive, and in particular to recognition when an articulatory combination occurs between words or syllables. Concerning a method for improving accuracy.

[Technology background]

Generally, when a plurality of words or syllables are uttered in succession, a deformation called articulatory combination occurs at the ends of adjacent words or syllables, especially when uttered rapidly. For example, when [aKa] and [aSa] are uttered consecutively, it becomes [aK aa Sa], and the two aas in the middle are combined and tend to become shorter. Therefore, there was a problem in that the matching accuracy with the prepared standard pattern decreased. A possible solution to this case would be to register in advance the changes due to articulatory combinations for all standard patterns, but most of the changes depend on the combination of standard patterns that cause articulatory combinations and the depth of the articulatory combinations. Therefore, it was difficult in practice.

[Object and structure of the invention]

An object of the present invention is to create a standard pattern of pseudo articulatory combination changes using a simple method in recognition of continuous speech input patterns including articulatory combinations, and to improve matching accuracy with input patterns.

To this end, the present invention has a structure in which acoustic features obtained by analyzing unknown input speech in which a plurality of words or syllables are successively uttered are stored in each part of an input pattern in advance. In a continuous speech recognition method that matches the required number of standard patterns of word/syllable sounds, if the ends of each adjacent standard pattern are separated by the boundary, the standard pattern of the separated part is By applying the interpolated pattern obtained from the ends of the standard pattern as It is characterized by outputting words or syllable sequences as recognition results.

[Embodiments of the invention]

The details of the present invention will be explained below with reference to the drawings.

FIG. 1 is an explanatory diagram of changes in articulation ratio in a continuous speech input pattern. The horizontal axis is the input pattern,
The vertical axis shows the standard pattern series. The adjacent section of width P (between m ₁ and l ₂ ) between partial patterns C ₁ and C ₂ in the input pattern is deformed and shortened by articulatory combination, and therefore, the input partial patterns C ₁ , C Standard patterns A and _B , which are prepared to be matched to pattern A and B, respectively, have overlapping ends, and pattern A
The terminal end P _eA of pattern B or the start end P _SB of pattern B is
In both cases, the articulatory supply deformation part m ₁ in the input pattern,
The matching characteristics for l ₂ deteriorate, causing increased ambiguity.

For this reason, in the present invention, as a standard pattern, P _SB and B _eB are removed in advance from the edge part of the pattern that is highly susceptible to deformation based on articulatory coupling, for example, standard pattern B in FIG. 1, and the width of only the central part is reduced. It uses a narrow pattern of
However, as a result, when pattern matching is performed, gaps are created between each standard pattern. This corresponds to the section Q transformed by articulatory combination in the input pattern. Therefore, a standard pattern for the articulatory combination section is created in advance in a pseudo manner. That is, 2
For all combinations of two narrow standard patterns, interpolation patterns are created that connect their ends, and during pattern matching, a matching interpolation pattern is selected for the gap between the narrow standard patterns. It is designed so that it fits into place.

FIG. 2 is an explanatory diagram of an interpolation pattern according to the present invention, where n and n ₂ are narrow standard patterns, and n _c is an interpolation pattern. As the interpolation method, an appropriate method such as linear interpolation or quadratic interpolation can be used. Further, the length of the interpolation pattern may be changed depending on (n ₁ , n ₂ ).

Next, an example of continuous speech recognition using such an interpolation pattern will be described.

As shown in Figure 3, the distance between the standard pattern n _i excluding both ends and the input partial pattern C _i (l, m) is D(l, m, n _i ), and the interpolation pattern n _i ' and the input The corresponding articulatory combination interval C _i ′(l′, m′) of the pattern
Let the distance between the two points be D _n (l', m', n _i -1, n _i ). In this case, if the minimum cumulative distance is S ₀ and l, m, l', and m' are generalized by l _i '+1, l _i , l _i-1 +1, and l _i ', respectively, it can be obtained by the following equation. It will be done.

S ₀ = min k _k 〓 _i-1 min min li, l′i, ni [D _n (l _i-1 +1, l _i ′, n _i −1, n _i ) + D
(l _i '+1, l _i , n _i )]...(1) Then, n _i (i=1, ₂ ,...,
Let k) be the recognition result.

FIG. 4 is a configuration diagram of the embodiment system, in which 1 is a standard pattern storage unit, 2 is a standard pattern storage unit, and 2 is a standard pattern storage unit.
Storage part of interpolation pattern created from n _i in preprocessing,
3 is a storage unit for the input pattern to be recognized, 4 is a distance calculation unit that calculates D _n +D in [ ] in equation (1), 5
is the minimum value calculation part of the distance calculation result, and 6 is the minimum cumulative distance S ₀ n _i (i=1, 2, 3, ..., k)
This is a similarity calculation unit that determines the Similarity calculation unit 6
The processing result is extracted as a recognition output.

Next, an example of the calculation procedure of equation (1) will be described. First, the minimum distance D^(l, m, n ₁ , n ₂ ) = min min j [D _n (l, j, n _i , n ₂ ) + D (j+1m, n ₂ )]...(2
) and remember it.

next, S (1, i, n) = D (1, i, n) ...(3) Solve the following recurrence formula using as the initial value.

S(k, i, n) = min min j, n' [S(k-1, j, n') + D^(j+1, i, n
', n)]...(4) At this time, at the same time B (k, i, n) = argmin j min min n' [S (k-1, j, n') + D^ (j+1, i, n' ,
n)]...(5) N(k-1, i, n) = argmin n' min min j[S(k-1, i, n'+D^(j+1, i, n', n)
]...Find and memorize (6).

The minimum cumulative distance is given by S ₀ = min k, nS (k, I, n) (7).

In order to obtain further recognition results, k ₀ = argmin k min nS (k, I, n) ...(8) N _k0 = argmin nS (k ₀ , I, n) ... (9), and then k=k ₀ i=I (10) as the initial value, N _k-1 = N (k-1, i, N _k )... (11) i and k are replaced as follows.

i←B (k, i, N _k ) k←k−1 (12) If k2, repeat from equation (11). N ₁ , N ₂ , . . . , Nk0 obtained in this way are the recognition results.

〔Effect of the invention〕

As described above, according to the present invention, the standard pattern is mainly composed of the central part of the pattern that is stable against articulatory combination, and the edge parts of the pattern that are likely to fluctuate due to articulatory combination are created by pseudo interpolation. By using this, it is possible to realize an efficient continuous speech recognition system with a relatively simple configuration.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram of conventional pattern matching when the input pattern includes an articulatory combination, FIG. 2 is an explanatory diagram of an interpolation pattern according to the present invention, and FIG. 3 is an explanatory diagram of pattern matching using an interpolation pattern. ,
FIG. 4 is a configuration diagram of a system according to an embodiment of the present invention. In the figure, 1 is a standard pattern storage section, 2 is an interpolation pattern storage section, 3 is an input pattern storage section, 4 is a distance calculation section, 5 is a minimum value calculation section, and 6 is a similarity calculation section.

Claims (1)

[Scope of Claims] 1. Preliminary information stored in each part of an input pattern representing acoustic features obtained by analyzing perceptual input speech in which a plurality of words or syllables are uttered in succession. In a continuous speech recognition method that matches the required number of standard patterns of word/syllable sounds, if the ends of each adjacent standard pattern are separated by the boundary during the matching, the standard pattern of the separated part is By applying the interpolation pattern obtained from the ends of the reference pattern, the degree of similarity with the input pattern is calculated, the standard pattern sequence with the maximum similarity is found, and the system corresponds to the obtained standard pattern sequence. A continuous speech recognition method characterized by outputting a word or syllable sequence as a recognition result. 2. In the continuous speech recognition method described in item 1 above, during matching, the partial similarity between each standard pattern and all input partial patterns of arbitrary length, and the end and start ends of all standard patterns are determined in advance. The method is characterized by calculating the interpolation pattern part similarities between all the interpolation patterns found from each part and all input partial patterns of arbitrary length, and then finding the optimal standard pattern sequence using dynamic programming. Continuous speech recognition method.