Synthesizing word baseforms used in speech recognition

• US 4,882,759 A
• Filed: 04/18/1986
• Issued: 11/21/1989
• Est. Priority Date: 04/18/1986
• Status: Expired due to Term

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Picheny, Michael A., Bahl, Lalit R., deSouza, Peter V., Mercer, Robert L.
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US06/853,525
• Time in Patent Office: 1,313 Days
• Field of Search: 381/36-43, 381/51-53, 364/513.5
• US Class Current: 704/243
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence

Abstract

Apparatus and method for synthesizing word baseforms for words not spoken during a training session, wherein each synthesized baseform represents a series of models from a first set of models, which include: (a) uttering speech during a training session and representing the uttered speech as a sequence of models from a second set of models; (b) for each of at least some of the second set models spoken in a given phonetic model context during the training session, storing a respective string of first set models; and (c) constructing a word baseform of first set models for a word not spoken during the training session, including the step of representing each piece of a word that corresponds to a second set model in a given context by the stored respective string, if any, corresponding thereto.

Claims

1. A method of synthesizing word baseforms for words wherein each synthesized baseform represents a series of models from a first set of models, each model of the first set corresponding to an output-related model wherein each output-related model correlates to an output generable by an acoustic processor, the method comprising the steps of:

(a) forming words as respective sequences of models from a second set of models, each model of the second set corresponding to a phonetic model;

(b) for a second set model occurring in a given context of second set models in step (a), storing a respective string of first set models;

(c) performing step (b) for each of at least one second set model; and

(d) constructing a word baseform of first set models for a word formed in step (a), including the step of representing each piece of a word that corresponds to a second set model in a given context by the stored respective string of first set models, if any, corresponding thereto.

2. A method of synthesizing word baseforms for words not spoken during a training session, wherein each synthesized baseform represents a series of output-related models and wherein each output-related model correlates to an output generatable by an acoustic processor, the method comprising the steps of:

(a) representing each of N words by a respective sequence of phonetic models, the positioning of a subject phonetic model relative to other phonetic models forming a phonetic context for the subject phonetic model;

(b) representing M words spoken during a training session by a series of output-related models, where the M words form a subset of the N words;

(c) for at least one subject word, aligning the phonetic models for the subject word against the output-related models for the subject word, the subject word having been spoken during the training session;

(d) from the alignment of output-related models and phonetic models, associating a string of output-related models with each of at least one respective phonetic model in a given context; and

(e) constructing an output-related model baseform for a word not spoken during the training session including the steps of (i) correlating a piece of said word not spoken during the training session to a phonetic model in a defined context;

(ii) determining if the phonetic model in said defined context corresponds to a similarly contexted phonetic model that has an associated string of output-related models; and

(iii) representing said word piece by said associated string.

3. The method of claim 2 comprising the further step of:

(f) identifying the context of a phonetic model based on the positioning of the phonetic model relative to n other phonetic models, where n is an integer greater than or equal to zero;

wherein the associated string of step (e)(iii) is the string of output-related models associated with the similarly contexted phonetic model having the highest n value.

4. The method of claim 3 wherein step (e) further includes the step of:

(iv) for said word not spoken during the training session, repeating steps (i), (ii), and (iii) for each piece thereof;

each word piece having a string of output-related models associated therewith.

5. The method of claim 4 wherein step (e) further includes the step of:

(v) concatenating the strings in the same order as the word pieces said strings represent to form a word baseform of output related models.

6. A method of synthesizing word baseforms for words not spoken during a training session, wherein each synthesized baseform represents a series of fenemic models and wherein each fenemic model correlates to a label output generatable by an acoustic processor, the method comprising the steps of:

(a) uttering known words during a training session, each known word uttered during the training session being formed of a known sequence of known word phonetic models;

(b) determining a context for a subject known word phonetic model based on the positioning of other phonetic models proximate to said subject phonetic model;

(c) repeating step (b) for one known word phonetic model after another as said subject phonetic model;

(d) representing at least some of the known words by a series of fenemic models;

(e) for a subject known word, aligning the phonetic models therefor against fenemic models representing said subject word;

(f) repeating step (e) for each of a plurality of subject words;

(g) associating a string of fenemic models with a respective known word phonetic model in a given context; and

(h) constructing a fenemic baseform for a new word not spoken during the training session including the steps of (i) correlating a piece of said new word not spoken during the training session with a new word phonetic model in a defined context;

(ii) determining if the new word phonetic model in said defined context corresponds to a similarly contexted known word phonetic model that has an associated string of fenemic models; and

(iii) representing said word piece by said associated string.

7. The method of claim 6 wherein step (g) comprises the step of:

(j) forming a string of fenemic models for a given known word phonetic model in a given context from multiple utterances during the training session of the given known word phonetic model in the given context.

8. The method of claim 7 wherein step (b) includes the step of:

(k) storing a context for a phonetic model based on the phonetic model occurring therebefore and the phonetic model occurring thereafter.

9. The method of claim 8 wherein step (h) further includes the steps of:

(l) determining whether a new word phonetic model has the same preceding phonetic model and the same succeeding phonetic model as a known word phonetic model; and

(m) representing the word piece corresponding to the new word phonetic model by the string of fenemic models for the known word phonetic model.

10. The method of claim 9 wherein step (h) further includes the steps of:

(n) determining whether a new word phonetic model has only the same preceding phonetic model as a known word phonetic model; and

(p) representing the word piece corresponding to the new word phonetic model by a leading portion of the string of fenemic models for the known word phonetic model.

11. The method of claim 10 wherein step (h) further includes the steps of:

(q) determining whether a new word phonetic model has only the same succeeding phonetic model as a known word phonetic model; and

(p) representing the word piece corresponding to the new word phonetic model by a trailing portion of the string of fenemic models for the known word phonetic model.

12. The method of claim 10 wherein step (h) includes the further step of:

(s) determining whether a new word phonetic model has neither the same preceding phonetic model nor the same succeeding phonetic model as a known word phonetic model; and

(t) representing the word piece corresponding to the new word phonetic model by a string of fenemic models for the known word phonetic model preceded or succeeded by any phonetic models.

13. In a speech recognition system which has a known phonetic baseform for each word, apparatus for synthesizing fenemic word baseforms for words not uttered during a training session, wherein each synthesized baseform represents a series of fenemic models based on label outputs generatable by an acoustic processor, the apparatus comprising:

(a) means for storing information identifying phonetic models and respective phonetic model contexts occurring during the training session;

(b) means for storing a respective string of fenemic models for a stored phonetic model in a given context occurring during the training session; and

(c) means for constructing a word baseform of fenemic models for a new word not spoken during the training session, which includes;

means for determining, for each new word phonetic model corresponding to a piece of the new word, a stored phonetic model and context which at least partially corresponds to the new word phonetic model in context; and

means for assigning to each piece of the new word at least a portion of the fenemic string stored for the determined stored phonetic model and context therefor.

14. A method of producing a synthesized word baseform for a first word not spoken during a training session, said synthesized baseform comprising a series of models from a first set of models, each model in the first set of models representing a unit of speech, each unit of speech having a size, said method comprising the steps of:

forming an intermediate word baseform of the first word, said intermediate word baseform comprising at least a second model from a second set of models, said second model having a context in the intermediate word baseform, each model in the second set representing a unit of speech different from the units of speech represented by the models of the first set of models, each unit of speech represented by the models of the second set having a size larger than the size of the largest unit of speech represented by a model of the first set;

correlating the second model, in its context in the intermediate word baseform, with a first series of models from the first set of models; and

replacing the second model in the intermediate word baseform with the first series of models to produce a synthesized word baseform.

15. A method as claimed in claim 14, characterized in that:

each model in the first set of model represents a fenemic unit of speech; and

each model in the second set of model represents a phonetic unit of speech.

16. A method as claimed in claim 14, characterized in that:

the intermediate word baseform comprises at least a third model from the second set of models, said third model preceding the second model in the intermediate word baseform; and

the step of correlating comprises correlating the second model, when it is preceded by the third model, with a first series of models from the first set of models.

17. A method as claimed in claim 14, characterized in that:

the intermediate word baseform comprises at least a fourth model from the second set of models, said fourth model following the second model in the intermediate word baseform; and

the step of correlating comprises correlating the second model, when it is followed by the fourth model, with a first series of models from the first set of models.

18. A method as claimed in claim 14, characterized in that:

the intermediate word baseform comprises at least third and fourth models from the second set of models, said third model preceding the second model in the intermediate word baseform, said fourth model following the second model in the intermediate word baseform; and

the step of correlating comprises correlating the second model, when it is preceded by the third model and when it is followed by the fourth model, with a first series of models from the first set of models.

19. A method as claimed in claim 14, characterized in that the step of correlating comprises the steps of:

forming an intermediate word baseform of a second word, said second word being different from the first word, said intermediate word baseform of the second word comprising at least the second model, said second model having a context in the intermediate word baseform of the second word which is the same as the context of the second model in the intermediate word baseform of the first word;

forming a training word baseform of the second word by uttering the second word, said training word baseform of the second word comprising a series of models from the first set of models; and

aligning the intermediate word baseform of the second word with the training word baseform of the second word to align the second model in the intermediate word baseform of the second word with a first series of models in the training word baseform of the second word.