JPH0120438B2 - - 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 particularly to recognition when a conjunctive combination occurs between words or syllables. This article relates to a method for improving accuracy.

[Technology background]

In general, when multiple words or syllables are uttered in succession, especially the faster they are uttered, the ends of adjacent words or syllables undergo transformations and shortenings called articulatory combinations, which result in preparedness. There was a problem that the matching accuracy with the existing standard pattern decreased. A possible solution to this case would be to register in advance changes due to articulatory combinations for all standard patterns, but there are many changes due to combinations of standard patterns that cause articulatory combinations and the depth of articulatory combinations. So much so that it was actually difficult.

[Object and structure of the invention]

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

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. When matching the required number of standard patterns of word sounds or syllables, if the edges of adjacent standard patterns overlap at the boundary, the overlapping part of each standard pattern is used as the standard pattern of the overlapping part. By using the averaged pattern obtained from the patterns of It is characterized by outputting a syllable sequence as a recognition result.

[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 articulatory combination changes and pattern matching in continuous speech input patterns. Subpatterns C ₁ and C ₂ in the input pattern
The adjacent sections (m ₁ , l ₂ ) of width Q are deformed and shortened by articulatory combination, and therefore, the standard patterns A, which should originally be matched with the input partial patterns C ₁ and C ₂ , respectively. B has a shape in which the ends overlap each other, 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 combination deformation section Q in the input pattern
The matching characteristics for (m ₁ , l ₂ ) deteriorate,
This causes an increase in ambiguity.

FIGS. 2a, b, and c are explanatory diagrams of the basic principle of the present invention. a is two adjacent standard patterns n ₁ ,
Indicates n ₂ . b is the standard pattern n ₁ , n ₂ , width P = 5
c is overlapped with width P=10. The pattern within the overlapping section is a pseudo-generated articulatory joint pattern. That is, as exemplified by u ₁ and u ₂ , two of all standard patterns (for example,
It is created by overlapping the ends of the combinations n ₁ , n ₂ ) and averaging them.

The averaging pattern can be created in various ways, such as by connecting two patterns smoothly, by connecting them with a straight line, or by simply performing arithmetic averaging at each position.

Next, an example of recognition processing using the above averaged pattern will be described.

As shown in Figure 3, the starting point of the standard pattern n _i
The distance between the input partial pattern C _{i- 1 (l, m) and the input partial pattern C i-1} (l, m) excluding the P i-1 frame and the terminal P _i frame is D(l, m, n _i , P _i-1 , P _i ).

Furthermore, the terminal P _i-1 frame of the standard pattern n _i-1 ,
The averaged pattern (P _i-1 frame) obtained from the starting point P _i-1 frame of the standard pattern n _i and the input partial pattern C _i ′ (l′, m′) corresponding to the articulatory combination section
_{Let the} distance from _the

Each l, m, l', m' as shown in the figure.
When generalized by l' _i+1 , l _i , l _i-1 +1, l _i ', the minimum cumulative distance S ₀ between patterns can be determined by the following equation.

S ₀ = min k _k 〓 ⁱ⁼¹ l ₁ , lmin ⁱ⁼¹ l ₁ , lmin i, ni, Pi, [Dm(l _i-1 +1, l _i ′, n _i-1 , n _i , _{Pi -1} )
+D(l _i '+1, l _i , n _i , P _i-1 , P _i )]...(1) n _i (i=1, 2,...k) which gives this S ₀ becomes 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.
3 is a storage unit for the averaging pattern created from n _i in preprocessing, 3 is a storage unit for the input pattern to be recognized, 4 is
In equation (1), the distance calculation unit calculates Dm+D in [ ], 5 is the minimum value calculation unit for the distance calculation result, and 6 is the minimum cumulative distance S ₀ n _i (i=1, 2, 3,...
k) is a similarity calculation unit that determines k). The processing result of the similarity calculation unit 6 is taken out as a recognition output.

Next, we will discuss the specific calculation procedure for equation (1).

First, D^(l ₁ , m, n ₁ , n ₂ , P ₁ , P ₂ ) = min min j [Dm(l, j, n ₁ , z ₂ , P ₁ )+D(j+1, m, n ₂ ,
P ₁ , P ₂ )]...Find (2). Next, with S(1, i, n, P) = D(1, i, n, o, P) ...(3) as the initial value, solve the following recurrence formula S(k, i, n, P) = min min j, n', P'{S((k-1, j, n', P') + D^(j
+1, i, n', n, P', P)}...(4) When calculating the above S(k, i, n, P), at the same time, B(k, i, n, P) = argmin j min min n', P'{S (k-1, j, n', P') + D^ (j+1,
i, n', n, P', P)}...(5) N(k-1, i, n, P) = argmin n' min min j', P'{S(k-1, j, n ′, P′) D^(j+1, i
, n', n, P', P)}...(6) P(k, i, n, P) = argmin P' min min j, n'{S(k-1, j, n', P' )D^(j+1,i
, n', n, P', P)}...(7) and memorize it.

As a result, the minimum cumulative distance S ₀ is given by S ₀ = min k, nS (k', I, n, o) (8).

Furthermore, in order to obtain the recognition result, k ₀ = argmin k min nS (k, I, n, o) ...(9) Nk ₀ = argmin nS (k ₀ , I, n, o) ... (10) , then use k=k ₀ , i=I, P=o...(11) as initial values to find N _k-1 =N(k-1, i, N _k , P)...(12), and then Replace i, P, and k simultaneously as follows.

i←B(k, i, _Nk , P) P←P(k, i, _Nk , P) k←k←1...(13) If k2, repeat from equation (12).

N ₁ , N ₂ , . . . , N _k0 obtained in this way are the recognition results.

〔Effect of the invention〕

As described above, according to the present invention, even if the input pattern includes an articulatory combination, accurate pattern matching can be performed relatively easily.

[Brief explanation of drawings]

Fig. 1 is an explanatory diagram of pattern matching of input patterns with articulatory combinations, Fig. 2 is an explanatory diagram of the averaging pattern based on the present invention, Fig. 3 is an explanatory diagram of recognition processing based on the present invention, and Fig. 4 is an explanatory diagram of the recognition processing based on the present invention. 1 is a configuration diagram of an example system. In the figure, 1 represents a standard pattern storage section, 2 an averaging pattern storage section, 3 an input pattern storage section, 4 a distance calculation section, 5 a minimum value calculation section, and 6 a similarity calculation section.

Claims (1)

[Claims] 1. Words previously stored in each part of an input pattern representing acoustic features obtained by analyzing unknown input speech in which a plurality of words or syllables are continuously uttered. When matching the required number of standard patterns of sounds or syllables, if the edges of adjacent standard patterns overlap at the boundary, the pattern of the overlapping part of each standard pattern is used as the standard pattern of the overlapping part. By using the averaged pattern obtained from each other, the similarity with the input pattern is calculated, the standard pattern sequence with the maximum similarity is found, and the word or syllable sequence corresponding to the obtained standard pattern sequence. A continuous speech recognition method characterized by outputting as a recognition result. 2. In the continuous speech recognition method described in item 1 above, when matching, all standard partial patterns whose starting and ending parts are removed within a predetermined length of each standard pattern and all input arbitrary lengths are used for matching. All averaged patterns of combinations of equal length and all parts of arbitrary length of the input, calculated from the partial similarity with the partial pattern and the end and start parts of all standard patterns within a predetermined length. A continuous speech recognition method characterized by calculating an averaged pattern part similarity with a pattern and then determining an optimal standard pattern sequence using dynamic programming.