System and method for generating a phonetic baseform for a word and using the generated baseform for speech recognition

• US 6,389,395 B1
• Filed: 04/04/1997
• Issued: 05/14/2002
• Est. Priority Date: 11/01/1994
• Status: Expired due to Term

• Current Assignee: Cisco Technology, Inc. (Cisco Systems, Inc.)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Ringland, Simon P.
• Primary Examiner(s): SMITS, TALIVALDIS IVARS
• Application Number:
  US08/817,072
• Time in Patent Office: 1,866 Days
• Field of Search: 704/250, 704/251, 704/256, 704/276, 704/241, 704/243, 704/244, 704/245, 704/270, 704/275, 704/242, 704/254
• US Class Current: 704/254
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

Out-of-vocabulary word models for a speech recognizer vocabulary are generated by forming phonemic transcriptions (phonetic baseforms) of user'"'"'s utterances in terms of existing reference phonemes by using a speech recognition algorithm to match input sub-word feature sample sequences to suitably-constrained allowable sequences of existing reference phoneme features. The resultant new-vocabulary-word phonetic baseform models are stored for subsequent speech recognition using the same recognition algorithm.

Claims

1. A method for generating a vocabulary for a speech recognition apparatus comprising the steps of:

receiving an input speech signal representing a word;

deriving feature samples from the received speech signal;

comparing the feature samples with allowable sequences of reference sub-word representations, at least one of said sub-word representations being capable of representing a sequence of more than one feature sample;

identifying, using an algorithm that is employed by a subsequent recognizer, the allowable sequence of reference sub-word representations that most closely resembles the received speech signal, and generating a coded representation from said identified representations; and

storing the generated coded representation of the word for subsequent recognition of another speech signal.

2. The method of claim 1, wherein the reference sub-word models are Hidden Markov Models.

3. A method for generating a vocabulary for a speech recognition apparatus comprising the steps of:

receiving an input speech signal representing a word;

deriving feature samples from the received speech signal;

comparing the feature samples with allowable sequences of reference sub-word representations, at least one of said sub-word representations being capable of representing a sequence of more than one feature sample;

identifying, using a Viterbi algorithm, the allowable sequence of reference sub-word representations that most closely resembles the received speech signal, and generating a coded representation from said identified representations; and

storing the generated coded representation of the word for subsequent recognition of another speech signal using a Viterbi algorithm.

4. The method according to claim 3, wherein all possible sequences of the reference sub-word representations are allowable.

5. The method according to claim 3, wherein the allowable sequences of sub-word representations are constrained to sequences that comprise sub-word representations that represent noise followed by sub-word representations that represent speech followed by sub-word representations that represent noise.

6. The method according to claim 3, wherein the step of identifying includes consideration of stored parameters each representing a transition probability of a sub-word representation following a previous sub-word representation.

7. The method according to claim 3, further comprising the step of generating a recognition network from one or more stored sub-word representations, said network representing allowable sequences of sub-word representations in the generated vocabulary.

8. The method according to claim 3, wherein the sub-word representations are statistical models.

9. The method according to claim 8, wherein the sub-word representations are Hidden Markov Models.

10. An apparatus for generating a vocabulary from an input speech signal, comprising:

a store containing a plurality of reference sub-word representations;

a feature deriver for receiving the input speech signal and operable to generate feature samples;

a recognizer connected to receive the generated feature samples, the recognizer having a vocabulary of allowable sequences of sub-word representations, at least one of said sub-word representations being capable of representing a sequence of more than one feature sample;

the recognizer being arranged in operation to employ a Viterbi algorithm to;

compare the received feature samples with the allowable sequences of reference sub-word representations; and

generate a coded representation by identifying an allowable sequence of reference sub-word representations that most closely resembles the input speech signal; and

a first store for storing the coded representation of the input speech signal for subsequent recognition of another speech signal using a Viterbi algorithm.

11. The apparatus according to claim 10, further comprising:

a recognizer arranged to compare input speech signals with coded representations in the first store and to output a signal indicative of recognition.

12. The apparatus according to claim 11, further comprising:

a second store of coded representations of words, said coded representations having been generated in a manner different from the coded representations stored in the first store.

13. The apparatus according to claim 12, wherein the coded representations of words identify a sequence of the reference sub-word representations.

14. The apparatus according to claim 10, wherein the vocabulary defines sequences of sub-word representations that comprise sub-word representations that represent noise followed by sub-word representations that represent speech followed by sub-word representations that represent noise.

15. The apparatus according to claim 10, wherein the vocabulary defines all possible sequences of sub-word representations.

16. The apparatus according to claim 10, wherein generation of the allowable sequence of reference sub-word representations that most closely resembles the received speech signal includes consideration of stored parameters each representing a transition probability of a sub-word representation following a previous sub-word representation.

17. The apparatus according to claim 10, wherein the sub-word representations are statistical models.

18. The apparatus according to claim 17, wherein the sub-word representations are Hidden Markov Models.

19. A method for generating a vocabulary for a speech recognition apparatus comprising the steps of:

receiving an input speech signal representing a word;

deriving feature samples from the received speech signal;

comparing the feature samples with allowable sequences of reference sub-word representations, at least one of said sub-word representations being capable of representing a sequence of more than one feature sample;

identifying, using a Viterbi algorithm which imposes transitional probabilities between each pair of sub-words, the allowable sequence of reference sub-word representations that most closely resembles the received speech signal, and generating a coded representation from said identified representations; and

storing the generated coded representation of the word for subsequent recognition of another speech signal using a Viterbi algorithm.

20. An apparatus for generating a vocabulary from an input speech signal, comprising:

a store containing a plurality of reference sub-word representations;

a feature deriver for receiving the input speech signal and operable to generate feature samples;

a recognizer connected to receive the generated feature samples, the recognizer having a vocabulary of allowable sequences of sub-word representations, at least one of said sub-word representations being capable of representing a sequence of more than one feature sample;

the recognizer being arranged in operation to employ a Viterbi algorithm to;

compare the received feature samples with the allowable sequences of reference sub-word representations; and

generate a coded representation by identifying an allowable sequence of reference sub-word representations that most closely resembles the input speech signal, taking into account transition probabilities between pairs of sub-words; and

a first store for storing the coded representation of the input speech signal for subsequent recognition of another speech signal using a Viterbi algorithm.