Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Patent
1996-06-18
2000-05-09
Hudspeth, David R.
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
382225, G10L 506, G10L 900
Patent
active
060616520
DESCRIPTION:
BRIEF SUMMARY
FIELD OF THE INVENTION
The present invention relates to a speech recognition apparatus such as a speech recognition device.
BACKGROUND OF THE INVENTION
There is a known method of using HMM (Hidden Markov Model) or DP (Dynamic Programming) matching for speech recognition. Both these methods are frequently used as basic technologies of speech recognition, an important problem in these systems for realizing expansion of vocabulary or continuous speech recognition and the like is to reduce the amount of calculation without deteriorating their functions. A method using vector quantization has already been proposed as a method of solving the problem. The present application concerns its improvement. Before entering the subject an explanation will first be given generally of HMM and DP matching and of how the technology of vector quantization is used.
HMM is considered to be a model for generating time-sequential signals in accordance with a stochastic property. In speech recognition using HMM, HMM r is provided in correspondence with a recognition unit r(=1, . . . ,R) such as a word, a syllable, a phoneme or the like (hereinafter, representatively word) to be recognized. When a vector series Y=(Y.sub.1,Y.sub.2, . . . ,Y.sub.T) (y.sub.t : a vector observed at a time point t) is observed, a degree of occurrence of Y is calculated from each HMM r and a word which corresponds to a HMM having the maximum degree is rendered the result of recognition.
FIG. 1 shows an example of HMM circle mark .smallcircle. designates a state of a system to be modeled by HMM, an arrow mark .fwdarw. designates a direction of transition of a state and q.sub.i designates a state i. Assume that a transition from a state i to a state j is caused by a probability a.sub.ij. The model is called a Markov chain when only a state and its transition probability are defined. By contrast, in HMM, a vector is assumed to generate on each state transition and .omega..sub.ij (y) is defined as a degree of occurrence of a vector y in accordance with a state transition q.sub.i .fwdarw.q.sub.j. There also are many cases expressed in .omega..sub.ij (y)=.omega..sub.ii (y) .omega..sub.i (y) or .omega..sub.ij (y)=.omega..sub.jj (y)=.omega..sub.j (y) in which y occurs not in accordance with state transition but with state. An explanation will be given in this application assuming that y occurs in accordance with state. Here, the structure of HMM, state transition probability and an occurrence probability of a vector are determined such that behavior of an object (speech pattern such as a word when used in speech recognition) to be modeled by HMM is explained as correctly as possible. FIG. 2 is an example of a constitution of HMM which is frequently used in speech recognition.
When HMM is defined, a degree L(Y.vertline..lambda.) whereby an observation vector series Y occurs from a model (designated by .lambda.) and can be calculated as follows. ##EQU1## where X=(X.sub.1,X.sub.2 . . . ,X.sub.T+1) designates a state series and .pi..sub.1 designates a probability of being in state i at t=1. In this model x.sub.t .epsilon.{1,2 . . . ,J,J+1) and x.sub.T+1 =J+1 is a final state. In the final state only a transition thereto occurs and there is no occurrence of vector.
HMM is grossly classified into a continuous type and a discrete type. In the continuous type, .omega..sub.i (y) is a continuous function such as a probability density function of y in which a degree of occurrence of y.sub.t is given as a value of .omega..sub.i (y) when y=y.sub.t. A parameter specifying .omega..sub.i (y) for each state i is defined and the degree of occurrence of y.sub.t at a state i is calculated by substituting y.sub.t into .omega..sub.i (y). For example, when .omega..sub.i (y) is given by a normal distribution of multiple dimensions, .SIGMA.i.
In the discrete type, an occurrence probability b.sub.im of a label m .epsilon.{1,2, . . . M} into which y.sub.t is to be transformed by vector quantization is stored in a table for each state i. The degree of occurrence of y.sub.t in the state
REFERENCES:
patent: 4570232 (1986-02-01), Shikano
patent: 4829577 (1989-05-01), Kuroda et al.
patent: 5040215 (1991-08-01), Amano
patent: 5050215 (1991-09-01), Nishimura
patent: 5058166 (1991-10-01), Ney et al.
patent: 5129002 (1992-07-01), Tsuboka
patent: 5263120 (1993-11-01), Bickel
patent: 5307444 (1994-04-01), Tsuboka
patent: 5608841 (1997-03-01), Tsuboka
E. Tsuboka et al., "On the Multiplication Type FVQ/HMM", Technical Report of IEICE, SP-93-27 Jun. 1993), pp. 25-31.
Japanese language search report dated Sep. 19, 1995.
English translation of Japanese language search report.
El Gamal A., et al, Using Simulated Annealing to Design Good Codes, IEEE Trans. Info. Theory vol. IT-33 No. 1, pp. 116-123, Jan. 1987.
Liu L., et al, Segment Matrix Vector Quantization and Fuzzy Logic for Isolated-word Speech Recognition, IEEE ISCS, pp. 152-156, Jun. 11, 1991.
Gersho, A., Gray R. M., Vector Quantization and Signal Compression, Kluwer, pp. 362-372, Dec. 1991.
Marques F., Kuo C-C, Classified Vector Quantization using Fuzzy Theory, IEEE ICFS, pp. 237-244, Mar. 1992.
Tsuboka E., Nakahashi J., On the Fuzzy Vector Quantization based Hidden Markov Model, ppl-637-640, Apr. 1994.
Cong L., et al, Combining Fuzzy Vector Quantization and Neural Network Classification for Robust Isolated Word Speech Recognition, Proceedings of ICCS, pp. 884-887, Nov. 1994.
Ho-Ping Tseng et al., "Fuzzy Vector Quantazation Applied to Hidden Markov Modeling", CH2396-0/87/0000-0641 IEEE, at 641-644 (1987).
Sugiyama, Masahide, "Automatic Language Recognition Using Acoustic Features," CH2977-7/91/0000-0813, IEEE, pp. 813-816 dated May 1991.
Nakahashi Jun-ichi
Tsuboka Eiichi
Hudspeth David R.
Matsushita Electric - Industrial Co., Ltd.
Sax Robert Louis
LandOfFree
Speech recognition apparatus does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Speech recognition apparatus, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Speech recognition apparatus will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1074561