Speech recognition system and method using a hidden markov model adapted to recognize a number of words and trained to recognize a greater number of phonetically dissimilar words.

• US 5,799,278 A
• Filed: 07/02/1996
• Issued: 08/25/1998
• Est. Priority Date: 09/15/1995
• Status: Expired due to Fees

• Current Assignee: Lenovo Singapore Pte Limited (Lenovo Group Ltd.)
• Original Assignee: International Business Machines Corporation
• Inventors: Cobbett, Michael, Pickering, John Brian
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Storm, Donald L.
• Application Number:
  US08/673,862
• Time in Patent Office: 784 Days
• Field of Search: 395/2.54, 395/2.53, 395/2.52, 395/2.65, 395/2.64, 395/2.86, 395/2.6, 395/2.51, 395/2.66, 395/2.44, 395/2.59, 395/759, 395/77, 395/23, 395/20, 704/232, 704/236, 704/200
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

A speech recognition system for discrete words uses a single Hidden Markov Model (HMM), which is nominally adapted to recognise N different isolated words, but which is trained to recognise M different words, where M>N. This is achieved by providing M sets of audio recordings, each set comprising multiple recordings of a respective one of said M words being spoken. Only N different labels are assigned to the M sets of audio recordings, so that at least one of the N labels has two or more sets of audio recordings assigned thereto. These two or more sets of audio recordings correspond to phonetically dissimilar words. The HMM is then trained by inputting each set of audio recordings and its assigned label. The HMM can effectively compensate for the phonetic variations between the different words assigned the same label, thereby avoiding the need to utilise a larger model (i.e., to use M labels).

Claims

1. A speech recognition system for discrete words, comprising:

interface means for receiving incoming voice signals;

processing means, operatively coupled to said interface means, for processing said incoming voice signals;

program means, responsive to said processed voice signals from said processing means, for performing speech recognition on said processed voice signals, said program means using a single Hidden Markov Model (HMM), said HMM nominally being adapted to recognise N different words, characterised in that said HMM is trained to recognise M different words, where M>

N, said M words being phonetically dissimilar from one another.

2. The speech recognition system of claim 1, in which M is greater than or equal to twice N.

3. The speech recognition system of claim 2, wherein said M words comprise N digits from a first language, and N corresponding digits from a second language.

4. The speech recognition system of claim 3, in which N is less than 100.

5. The speech recognition system of claim 2, in which N is less than 100.

6. The speech recognition system of claim 1, in which the number of states per nominal word is S, whereby the HMM has a total of N×

S states.

7. The speech recognition system of claim 6, in which the HMM has N allowable final states.

8. The speech recognition system of claim 7, in which S is less than 30.

9. The speech recognition system of claim 6, in which S is less than 30.

10. The speech recognition system of claim 1, in which N is less than 100.

11. A computer implemented method for training a speech recognition system for discrete words, comprising the steps of:

providing a single Hidden Markov Model (HMM);

providing M sets of audio recordings, each set including multiple recordings of a respective one of said M words being spoken;

assigning N labels to the M sets of audio recordings, such that at least one of the N labels has two or more sets of audio recordings assigned thereto, said two or more sets of audio recordings corresponding to phonetically dissimilar words;

inputting each of said sets of audio recordings and assigned N labels into said speech recognition system;

determining a training path through said HMM of said system for each of said inputted audio recordings; and

storing said determined training path for each of said audio recordings together with said N label assigned to each of said audio recordings, whereby speech recognition of a word can be performed by determining said training path most likely to output said word to be recognised and then equating said word to be recognised with said N label associated with said most likely determined path.

12. The method of claim 11, in which M is greater than or equal to twice N, and each label has two or more sets of audio recordings assigned thereto.

13. The method of claim 12, wherein said M words comprise N digits from a first language, and N corresponding digits from a second language.