JPS6335997B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention samples an input speech signal at regular intervals, converts it into a phoneme sequence, and then compares it with each word in a word dictionary registered in advance using phoneme symbols, and selects the word with the highest degree of similarity as a recognition result. In the speech recognition device, the purpose is to increase the speed of comparison with the word dictionary by merging the converted phoneme sequences. FIG. 1 is a block diagram showing the conventional structure of a speech recognition device that once decomposes an input speech signal into a series of phonemes and then performs word recognition. Reference numeral 1 denotes an audio signal input terminal, and 2 represents a frequency analysis section that includes, for example, an n-channel filter bank and performs frequency analysis of the input audio signal. That is, with respect to the input audio signal, a voltage corresponding to the magnitude of the output of each band filter constituting the filter bank is outputted as the magnitude of each frequency component. For example, if you sample this every 10msec,
The audio signal is converted into an n-dimensional vector sequence every 10 msec. (Here each
(10 msec is called a frame) 3 is a standard pattern storage unit that stores standard patterns of phonemes, and each phoneme is stored as an n-dimensional vector obtained as a result of frequency analysis of each phoneme using the filter bank. . 4 is a phoneme recognition unit that calculates which phoneme pattern in the standard pattern storage unit 3 each of the n-dimensional vector series output from the frequency analysis unit 2 is closest to, and corresponds to the closest standard pattern. Outputs the phoneme that is recognized as a recognition result. At this point, the input speech has been converted into a phoneme sequence. Reference numeral 5 denotes a word dictionary, in which each word constituting the vocabulary to be recognized is registered in advance in the form of a combination of phonemes, roughly written in Roman letters. 6
is a word recognition unit that compares the phoneme sequence obtained by the phoneme recognition unit 4 with each word in the word dictionary 5 based on the degree of similarity between phonemes determined experimentally or theoretically between each phoneme in advance. A comparison is made and the word with the highest degree of similarity is output as the recognition result. 7 is an output terminal for the determination result. In this case, the phoneme sequence recognized by the phoneme recognition unit 4 contains many mistakes, so instead of making each of the input n-dimensional vectors correspond to a unique phoneme, the first candidate phoneme , the second candidate phoneme, and the first candidate phoneme are output as a reliability set. Figure 2 shows an example of this. That is, for the i-th vector Vi, A _i is the first candidate phoneme, and B _i is the second candidate phoneme.
Candidate phoneme, r _i is the reliability of the first candidate. To determine these phonemes and reliability, calculate the distance between each standard pattern vector and Vi, select the standard pattern with the shortest distance as the first candidate, then choose the standard pattern with the shortest distance as the second candidate, and select the standard pattern with the shortest distance as the second candidate. When the distance is d _i1 and the latter distance is d _i2 , the reliability r _i is r _i = d _i2 /
It is obtained by (d _i1 + d _i2 ). The distance between the standard pattern vectors Ai and Bi for the phonemes A _i and B _i and the input vector V _i can be defined as the distance between normal vectors. For example, if this distance is defined as Euclidean distance, then Ai = (A _i1 , A _i2 , ..., A _io ) Bi = (B _i1 , B _i2 , ..., B _io ) Vi = (V _i1 , V _i2 , ..., V _io ) become. In this way, when the certainty of Ai and Bi is about the same, ri=0.5, and as the certainty of Ai increases, ri approaches 1. In order to compare the phoneme series obtained in this way with each word in the word dictionary, it is necessary to define the degree of similarity between each of the phoneme series and the phonemes that make up each word. . Similarity between phonemes For example, the similarity between phonemes P and Q is the distance between the two phonemes obtained by statistically processing a large amount of data collected as n-dimensional vectors corresponding to phonemes P and Q. It is obtained by linear transformation of . This is defined as similarity S ₀ (P, Q). At this time,
The degree of similarity S(i, j) between the kth word W _k and the jth phoneme D _j , which is the target of the comparison between the i-th phoneme set of the phoneme series and the word dictionary, is, for example, S(i , j)=riS ₀ (A _i , D _j ) + (1−ri) S ₀ (B _i , D _j ). The degree of similarity between the input phoneme sequence and the word W _k can be determined by well-known dynamic programming on a lattice graph with the vertical axis j and the horizontal axis i based on this S(i, j). In the above recognition, the output phoneme sequence of the phoneme recognition unit 4 is not input as is to the word recognition unit 6, but is usually merged to reduce the number of phonemes. This is done in order to increase the recognition speed, as it would be inefficient to directly check with a word dictionary because the same phoneme is often consecutive if the sampling is about 10 msec, and there is a high degree of redundancy. The phoneme recognition results are often erroneous in unstable areas such as transitions between phonemes, and this is done to eliminate this. The present invention provides a speech recognition device equipped with a new merging method in the above recognition method. As explained in Figure 2, if the first candidate phoneme of the i-th frame is A _i , the second candidate phoneme is B _i , and the reliability of A _i is ri, then the reliability of B _i is 1-ri. . Now, the jth
The phoneme sequence from the frame to the j+kth frame is merged, and the first candidate phoneme Fl and the second candidate phoneme are
We will explain how to obtain Sl and reliability Rl. The phonemes included in this range are A _j , A _(j+1) , ...,
A _(j+k) , B _j , B _(j+1) , ..., B _(j+k) , and each has the reliability rj, r(j+1),
..., r(j+k), 1-rj, 1-r(j+1),...
..., 1-r(j+k) are attached. Assume that there are m types of different phonemes, and let them be X ₁ , X ₂ , ..., X _n , then the above A _j ,
..., B _(j+k) will be included in any one of X ₁ , ..., X _n . At this time, A _j ,...,
Among B _(j+k) , let ui be the sum of the reliabilities corresponding to those included in X _i . If the largest one of u ₁ , u ₂ , ..., u _n is uλ and the second largest one is uμ, then the first candidate phoneme Fl=Xλ, the second candidate phoneme Sl=Xμ,
The reliability can be given as Rl=uλ/(uλ+uμ). FIG. 3 shows the configuration of a speech recognition device incorporating the merging method according to the present invention, in which 1 to 7 perform the same operations as the conventional example, and 8 is a phoneme string modification that performs the merging according to the present invention. Department. As for the range of merging, experimentally fixing it in units of three frames (that is, m=3) is the simplest method and is relatively effective. As an example, we will discuss the case of merging frames from i-1 to i+1 (when k=2). In FIG. 2, the reliability of each phoneme included in this frame is as shown in the table below.

【table】

[Table] For example, A _(i-1) , A _i , B _(i+1) are phonemes X ₁ ,
If B _(i-1) , B _i is the phoneme X ₂ and A _(i+1) is the phoneme X ₃ (if m = 3), then u ₁ = r (i-1) + ri + (1-r _(i+1) ) u ₂ = (1-r _(i-1) ) + (1-ri) u ₃ = r _(i+1) . Therefore, if u ₁ ≧u ₂ ≧u ₃ , then uλ=u ₁ and uμ=u ₂ , so the merged results are Fl=X ₁ , Sl=X ₂ , Rl=u1/(u ₁ + u ₂ ). As described above, according to the present invention, it is possible to rationally merge frames based on the number of identical phonemes included in the range to be merged and their reliability, that is, to reduce redundancy and create a word dictionary. This makes it possible to improve recognition speed and accuracy more efficiently than before.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device that performs phoneme analysis, Fig. 2 is a diagram explaining a phoneme sequence obtained as a result of phoneme recognition, and Fig. 3 shows an embodiment of the speech recognition device of the present invention. It is a block diagram. 2...Frequency analysis unit, 3...Standard pattern storage unit, 4...Phonological recognition unit, 5...Word dictionary, 6...
Word recognition section, 8...Phone sequence correction section.

Claims (1)

[Claims] 1 A means for converting an input speech signal into a series of feature vectors, and a means for converting each feature vector constituting the vector series into a first candidate phoneme, a second candidate phoneme, and a means for converting the first candidate phoneme and the second candidate phoneme with the same certainty. a phoneme recognition means that converts into a set of reliability of the first candidate phoneme, which is defined as 0.5 when the first candidate phoneme has a certainty, and approaches 1 as the certainty of the first candidate phoneme increases; means for merging several successive sets of a first candidate phoneme, a second candidate phoneme, and a reliability set of the first candidate phoneme, and an output sequence of the merging means and each word constituting the recognition vocabulary. In a speech recognition device that compares and collates a word with each word in a word dictionary expressed as a combination of phonemes and outputs the word with the highest degree of similarity as a recognition result, the merging means includes the i-th number of the output series of the phoneme recognition means. Let A _i be the first candidate phoneme of the set of
When the second candidate phoneme is B _i , the reliability of the first candidate phoneme is ri, and the reliability of the second candidate phoneme is 1−ri,
When merging sets from the j-th set to the j+k-th set, the m types of phonemes included therein X ₁ ,
For X ₂ , ..., X _n , the above A _j , A _j +1, ...,
Among A _j +k, B _j , B _j +1, ..., B _{j + k} , phoneme
Let u _i be the sum of the reliability corresponding to each thing equal to X _i , and let the maximum of u ₁ , u ₂ , ..., u _n be
If uλ is the next largest one, then the first candidate phoneme after merging is Xλ, and the second candidate phoneme is
A speech recognition device characterized in that it has means for setting the reliability of the first candidate phoneme to uλ/(uλ+uμ).