Fully expanded context-dependent networks for speech recognition

• US 6,574,597 B1
• Filed: 02/11/2000
• Issued: 06/03/2003
• Est. Priority Date: 05/08/1998
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Riley, Michael Dennis, Mohri, Mehryar
• Primary Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US09/502,501
• Time in Patent Office: 1,208 Days
• Field of Search: 704/255, 704/256, 704/231, 704/232, 704/242, 704/243, 704/250, 704/251, 704/257
• US Class Current: 704/251
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

A large vocabulary speech recognizer including a combined weighted network of transducers reflecting fully expanded context-dependent modeling of pronunciations and language that can be used with a single-pass Viterbi or other coder based on sequences of labels provided by feature analysis of input speech.

Claims

1. A method for making a combined weighted speech transducer for a large-vocabulary context-dependent speech recognizer based on signals representing:

(i) the inverse, C<sup>−

1</sup>, of a context-dependency transducer;

(ii) a word pronunciation transducer, L; and

(iii) a language model transducer, G;

the method comprising the steps of

generating signals representing transducer C, the inverse of a determinized version of transducer C<sup>−

1</sup>,

generating signals representing transducer P′

, a determinized version of the composition L′

BG′

, which composition is a composition of disambiguated versions of each of said transducers L and G,

generating signals representing a transducer P, a minimized version of transducer P′

, and

generating signals representing said combined speech transducer as the composition C B P.

2. The method of claim 1 wherein said step of generating signals representing transducer C comprises the steps of generating signals representing a determinized version of C⁻

1, and generating signals representing the inverse of said determinized version of C<sup>−

1</sup>.

3. The method of claim 1 wherein said step of generating signals representing P′

comprises

generating signals representing a transducer L′

, a disambiguated version of L,

generating signals representing transducer G′

, a disambiguated version of G,

generating signals representing a transducer L′

BG′

that is a determinized version of the composition of L′ and

G′

.

4. The method of claim 3, wherein said step of generating signals representing L′

comprises labeling with auxiliary labels those paths in L that map input strings to outputs in excess of a first output.

5. The method of claim 3, wherein said step of generating signals representing G′

comprises labeling with auxiliary labels those paths that map input strings to context sequences in excess of a first context sequence.

6. The method of claim 4, wherein said step of generating signals representing G′

comprises labeling with auxiliary labels those paths that map input strings to context sequences in excess of a first context sequence.

7. The method of claim 6, where in said step of generating signals representing P comprises steps of

modifying said transducer P′

by replacing said auxiliary labels by ε

, and removing ε

-arcs in said modified version of P′

.

8. The method of claim 1 wherein said language model, G, is an n-gram model, where n is a positive integer.

9. The method of claim 8 wherein n=2.

10. The method of claim 8 wherein n=3.

11. The method of claim 1 wherein said context dependency transducer, said inverse transducer C⁻

1, the determinized version of C<sup>−

1</sup> and the inverse of the determinized version of C<sup>−

1</sup> are cross-word context transducers.

12. The method of claim 1 wherein said combined weighted transducer is fully expanded.

13. The method of claim 1 wherein G is a weighted transducer.

14. The method of claim 1 wherein L is a weighted transducer.

15. The method of claim 1 wherein G and L are weighted transducers.

16. A combined weighted speech transducer for use in a large-vocabulary context-dependent speech recognizer, said transducer stored in a memory system and being based on signals representing:

(i) the inverse, C<sup>−

1</sup>, of a context-dependency transducer;

(ii) a word pronunciation transducer, L; and

(iii) a language model transducer, G;

said transducer comprising

signals representing C B P, the composition of transducers C and P, where

transducer C comprises signals representing the inverse of a determinized version of transducer C<sup>−

1</sup>,

transfer P comprises signals representing a minimized version of a transducer P′

, where transducer P′

is a determinized version of the composition L′

BG′

, of disambiguated versions of each of said transducers L and G.

17. A large-vocabulary, context-dependent speech recognizer comprising

a. a feature extractor for extracting features of input speech signals and applying sequences of one or more labels to said features,

b. a combined weighted speech transducer for use in a speech recognizer, said transducer being stored in a memory system and being based on signals representing:

(i) the inverse, C<sup>−

1</sup>, of a context-dependency transducer;

(ii) a word pronunciation transducer, L; and

(iii) a language model transducer, G;

said combined speech transducer comprising signals representing C B P, the composition of transducers C and P, where

transducer C comprises signals representing the inverse of a determinized version of transducer C<sup>−

1</sup>, and

transducer P comprises signals representing a minimized version of a transducer P′

, where transducer P′

is a determinized version of the composition L′

BG′

, of disambiguated versions of each of said transducers L and G, and

c. a decoder for outputting decisions about said input speech signals based on said sequences of labels and said combined speech transducer.

18. The speech recognizer of claim 17 wherein said decoder is a single-pass decoder.

19. The speech recognizer of claim 17 wherein said decoder is a Viterbi decoder.

20. The speech recognizer of claim 19 wherein said combined weighted speech transducer is fully expanded.