Large vocabulary connected speech recognition system and method of language representation using evolutional grammar to represent context free grammars

• US 5,907,634 A
• Filed: 04/25/1997
• Issued: 05/25/1999
• Est. Priority Date: 01/21/1994
• Status: Expired due to Term

• Current Assignee: AT&T Corporation (AT&T, Inc.)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Brown, Michael Kenneth, Glinski, Stephen Charles
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): Tadayon, Bijan
• Application Number:
  US08/846,014
• Time in Patent Office: 760 Days
• Field of Search: 704/252, 382/155, 382/156, 382/173, 382/181, 382/185, 382/186, 382/187, 382/188, 382/189, 382/190, 382/192, 382/195, 382/199, 382/201, 382/202, 382/203, 382/209, 382/217, 382/224 , 382/226, 382/227, 382/228, 382/229, 382/230, 382/231, 382/232, 382/254, 382/266, 382/270, 382/191, 382/193, 382/194, 382/196, 382/197, 382/205, 382/210, 382/215 , 382/216, 382/218, 382/220, 382/271, 382/272, 382/273, 382/172, 395/2.55, 395/2.5, 395/2, 395/2.09, 395/2.61, 395/2.42, 395/2.64, 395/2.52
• US Class Current: 382/226
• CPC Class Codes: G06F 18/29 Graphical models, e.g. Bayesian networks
• CPC Class Codes: G06V 30/347 Sampling; Contour coding; Stroke extraction
• CPC Class Codes: G06V 30/373 using a special pattern or subpattern alphabet
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

A method of recognizing speech input selectively creates and maintains grammar representations of the speech input in essentially real time. Speech input frames are received by a speech recognition system. Grammar representations are created for each speech frame and a probability score is derived for the representations indicating the probability of the accuracy of the representations to the speech input. Representations having a probability score below a predetermined threshold are not maintained. Those grammar representations having probability scores above the predetermined threshold are maintained. As more speech frames are received by the system, additional grammar representations are created and the probability scores are updated. When the entire speech input has been received, the chain of grammar representations having the highest probability score is identified as the speech input.

Claims

1. A method for use in a grammar-based recognition system wherein the grammar is represented by a grammar network comprised of arcs interconnecting nodes, the arcs being representative of recognition models, each arc having a source node and a destination node, and in which feature scores are generated by a processor in the system and input to the grammar network and resulting cumulative hypothesis scores are propagated through the models to produce cumulative hypothesis scores at various ones of the nodes, the method comprising the step of:

allocating resources in the processor to implement a recognition model, currently absent from the grammar network, for an individual one of the arcs in response to the production, at its source node, of a cumulative hypothesis score meeting a predetermined criterion, the arcs including at least one terminal arc representative of a predefined recognition model and at least one non-terminal arc representative of a predefined grammar sub-network comprising a plurality of non-terminal arcs, the at least one terminal arc and the at least one non-terminal arc being arranged in a predetermined order in the grammar network.

2. The method of claim 1 comprising the further step of

ceasing to propagate a hypothesis score through an individual one of said recognition models if cumulative hypothesis scores within said individual model satisfy a predetermined turn-off criterion.

3. The method of claim 2 wherein said turn-off criterion is that each of the cumulative hypothesis scores within said individual model fall below a predetermined turn-off threshold after having exceeded it.

4. The method of claim 3 comprising the further step of releasing memory used by said individual recognition model back to said system for use in later resource allocation to implement the recognition models for ones of said arcs other than said individual one of the arcs.

5. The method of claim 1 wherein each recognition model includes at least one Hidden Markov Model (HMM).