Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Patent
1997-09-24
2000-02-08
Hudspeth, David R.
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
704251, 704255, G10L 506, G10L 900
Patent
active
060236779
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
The invention relates to a speech recognition method and more particularly to a method of the type in which sentences that have been put together from words of a given vocabulary are recognized, wherein a limited number of permissible sentences and an N-gram speech model into which the syntax of the permissible sentences is integrated are predetermined.
In the recognition of connected speech, which permits any combination of all words, the error rate increases considerably compared to the recognition of individual words. To counteract this, knowledge on permissible word sequences, for example, can be stored in so-called speech models and can be used in the recognition. As a result, the number of permissible sentences can be limited considerably.
Usually, speech models are defined as N-gram models, with N being identified as the depth of the model and indicating the number of words following one another within a word sequence which are considered in the actual evaluation of a word sequence hypothesis. The recognition process becomes rapidly more complex as N increases; therefore, the particularly simple bigram model with N=2 is preferred which only considers combinations of two words. The speech models can be further simplified if words, which occur in the same context but which do not necessarily have to have the same meaning, are combined in word groups (e. g., all weekdays). Instead of individual word transitions, the speech models can consider the transition from one word group to another.
In Informatik Forsch. Entw. [Informatics Research Development] (1992) 7, p. 83-97, basic problems of the automatic recognition of flowing language are dealt with in detail and approaches for solving problems are described from the point of view of statistical decision theory. The focus is on the stochastic modelling of knowledge sources for acoustics and linguistics, e. g., in the form of phoneme models, pronunciation dictionary and speech model.
From "The HARPY Speach Understanding System" in Readings in Speech Recognition, 1990, Morgan Kaufmann Publishers Inc., a speech recognition system is known which has a greatly limited number of permissible sentences. The syntactic and semantic constraints determining permissibility can be formulated in grammar equations and can be represented as a graph. A few simplifications are introduced in order to get from the grammar definition, which is complete but involves great processing complexity, to a compact speech model with reasonable processing complexity.
But sometimes such simplifications are only possible if it is accepted for the speech model that nonpermissible word sequences in the original grammar definition appear as being permissible again. Finally, in the HARPY system, the words are replaced by their phonetic definitions and, in this manner, a phonetic model for a complete sentence recognizer is created.
SUMMARY OF THE INVENTION
It is the object of the present invention to propose a method for speech recognition which has a high recognition rate while the processing complexity is low.
According to the invention, a speech recognition method of the type mentioned at the outset is characterized in that for words which recur in the number of permissible sentences in different syntactical positions, recurrent specimens that can be distinguished from one another are predetermined in the speech model with the syntactical constraints that are valid for the respective position, and in that, through continuous consideration of the syntactical constraints of the speech model during the recognition process in progress, only the matching of an actual speech signal with permissible word sequences is examiner.
By way of the distinguishable identification of words which are recurring in the grammar of the number of permissible sentences, the invention permits the implicit and complete detection in the speech model of the permissible predecessors of a specific word in a defined sentence position without having to first explicitly store all permissibl
REFERENCES:
patent: 4277644 (1981-07-01), Levinson et al.
patent: 4882757 (1989-11-01), Fisher
patent: 5467425 (1995-11-01), Lau
patent: 5787230 (1998-07-01), Lee
Rabiner, L R, Juang, B-H, Fundamentals of Speech Recognition, AT&T/Prentice Hall, pp. 442-3, 448, 450-1, Jan. 1993.
Bruce Lowerre et al, "The Harpy Speech Understanding System", Trends in Speech Recognition, Speech Communications Research Laboratory and University of Southern California (1980), pp. 340-360.
Hermann Ney et al, "On structuring probabilistic dependencies in stochastic language modelling", 8300 Computer Speech & Language, vol. 8, No. 1, (Jan., 1994), pp. 1-38.
Hermann Ney, "Automatische Spracherkennung: Architektur und Suchstrategie aus statistischer Sicht", Informatics Research Development, vol. 7 (1992), pp. 83-97.
Sei-ichi Nakagawa, "Speaker-independent continuous-speech recognition by phoneme-based word spotting and time-synchronous context-free parsing", 8300 Computer Speech & Language, vol. 3, No. 3 (Jul., 1989), pp. 277-299.
Class Fritz
Kaltenmeier Alfred
Kilian Ute
Regel-Brietzmann Peter
Daimler - Benz AG
Hudspeth David R.
Sax Robert Louis
Spencer George H.
Wood Allen
LandOfFree
Speech recognition method 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 method, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Speech recognition method will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1688350