Speech recognition system for determining a recognition result at an intermediate state of processing

• US 5,875,425 A
• Filed: 12/23/1996
• Issued: 02/23/1999
• Est. Priority Date: 12/27/1995
• Status: Expired due to Fees

• Current Assignee: KDDI Corporation
• Original Assignee: Kokusai Denshi Denwa Company Limited (KDDI Corporation)
• Inventors: Yato, Fumihiro, Yamamoto, Seiichi, Nakamura, Makoto, Inoue, Naomi
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/772,987
• Time in Patent Office: 792 Days
• Field of Search: 704/245, 704/252, 704/241, 704/253, 704/244, 704/277, 704/238, 704/243, 704/231
• US Class Current: 704/231
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

A speech recognition system for recognizing a system user'"'"'s speech can shorten a recognition period by reducing the amount of necessary calculations without deteriorating the accuracy rate of recognition. The speech recognition system successively calculates statistical probabilities of acoustic models, outputs a one sentence recognition result corresponding to acoustic models having the highest reliability when the one sentence is detected and stops the following calculations.

Claims

1. A speech recognition method in a speech recognition system having a language model register for registering a plurality of language models and wherein a plurality of series of acoustic models are previously registered, comprising the steps of:

calculating statistical probabilities of a plurality of series of acoustic models;

selecting a series of acoustic models having the highest probability;

detecting sentences, formed by at least one word, stored in a language model register that correspond to the selected series of acoustic models;

determining a recognition result if only one sentence is detected in said language model register.

2. A speech recognition method as claimed in claim 1, wherein only words which are previously registered in said language model register can form the sentences stored in said language model register.

3. A speech recognition method as claimed in claim 2, wherein only a series of words which is grammatically correct can form the sentences stored in said language model register.

4. A speech recognition method as claimed in claim 1, wherein said language model register stores a plurality of sentences which previously have been spoken by a system user.

5. A speech recognition method as claimed in claim 1, wherein said plurality of series of acoustic models are formed by learning a time series of a plurality of previously formed acoustic models.

6. A speech recognition method as claimed in claim 1, wherein said only one sentence constantly remains in the language model register from a time at which said only one sentence is initially detected until a time at which the recognition result is determined.

7. A speech recognition method as claimed in claim 1, wherein said only one sentence constantly remains in the language model register from a frame at which said only one sentence is initially detected until a frame at which the recognition result is determined.

8. A speech recognition method, comprising the steps of:

calculating statistical probabilities of correspondence between each of a plurality of acoustic models stored in an acoustic model register and inputted voice sounds;

selecting a series of acoustic models having the highest probability of correspondence to the inputted voice sounds;

detecting sentences, formed by at least one word, stored in a language model register, said detected sentences corresponding to the selected series of acoustic models and following acoustic models capable of being appended to the selected series of acoustic models;

outputting a recognition result when only one sentence is detected in said language model register.

9. A speech recognition method as claimed in claim 8, wherein only words which are previously registered in said language model register can form the sentences stored in said language model register.

10. A speech recognition method as claimed in claim 8, wherein only a series of words which is grammatically correct can form the sentences stored in said language model register.

11. A speech recognition method as claimed in claim 8, wherein said language model register stores a plurality of sentences which previously have been spoken by a system user.

12. A speech recognition method as claimed in claim 8, wherein said series of acoustic models stored in said acoustic model register are formed by learning a time series of a plurality of previously formed acoustic models.

13. A speech recognition method for outputting a recognition result prior to detecting the end of an inputted word, comprising the steps of:

selecting a series of acoustic models having the highest probability of representing a beginning portion of an inputted word; and

outputting a recognition result when only one following acoustic model can be appended to the selected series of acoustic models to correspond to a grammatically correct sentence, formed by at least one word, stored in a language model register.

14. A speech recognition method which reduces the number of calculations necessary to recognize inputted voice sounds comprising the steps of:

detecting a beginning portion of an inputted word; and

outputting a recognition result when only one sentence, formed by at least one word, is detected in a language model register, wherein

completing portions are appended to said beginning portion to form sentences stored in said language model register,

only words which are previously registered in said language model register can form sentences stored in said language model register, and

only a series of words which is grammatically correct can form sentences stored in said language model register.

15. A speech recognition method, comprising the steps of:

selecting a beginning series of acoustic models having the highest probability of representing inputted voice sounds;

selecting a following series of acoustic model capable of being appended to the beginning selected series of acoustic model to form a grammatically correct sentence, formed by at least one word, stored in a language model register; and

outputting a recognition result when only one following series of acoustic models remains.