Speech recognition by acoustic/phonetic system and technique

• US 4,852,180 A
• Filed: 04/03/1987
• Issued: 07/25/1989
• Est. Priority Date: 04/03/1987
• Status: Expired due to Term

• Current Assignee: American Telephone & Telegraph Company (AT&T, Inc.), Bell Telephone Laboratories, Inc. (Nokia Corporation)
• Original Assignee: American Telephone & Telegraph Company (AT&T, Inc.)
• Inventors: Levinson, Stephen E.
• Primary Examiner(s): NOT, DEFINED
• Assistant Examiner(s): NOT, DEFINED
• Application Number:
  US07/034,467
• Time in Patent Office: 844 Days
• Field of Search: 381/41-43, 381/44-45, 364/513, 364/513.5
• US Class Current: 704/256.4
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence

Abstract

A speech recognition system and technique of the acoustic/phonetic type is made speaker-independent and capable of continuous speech recognition during fluent discourse by a combination of techniques which include, inter alia, using a so-called continuously-variable-duration hidden Markov vodel in identifying word segments, i.e., phonetic units, and developing proposed phonetic sequences by a durationally-responsive recursion before any lexical access is attempted. Lexical access is facilitated by the phonetic transcriptions provided by the durationally-responsive recursion; and the resulting array of word candidates facilitates the subsequent alignment of the word candidates with the acoustic feature signals. A separate step is used for aligning the members of the candidate word arrays with the acoustic feature signals representative of the corresponding portion of the utterance. Any residual work selection ambiguities are then more readily resolved, regardless of the ultimate sentence selection technique employed.

Claims

1. A method for the recognition of speech, of the type including the steps of

storing signals representing a model of the language to be recognized, said model being of the state-transitional type, each state being uniquely identified with a phonetic unit,

each state having associated with it a portion of a transition matrix which describes which states can follow it and with what probability,

each state having associated with it an observational density function assigning to each set of speech feature signals that may be observed in fluent speech a likelihood of being observed in association with that state,

each state having associated with it a durational density function assigning to each duration it may have a likelihood of occurrence in fluent speech;

storing signals representing lexical candidates, said lexical candidates being assemblages of phonetic units of the language in association with partial phonetic information of the type found in dictionaries;

sequentially converting successive time frame portions of an utterance into signals representing respective sets of acoustic feature signals representative of the portions; and

accessing the stored model and stored lexical candidates to obtain signals which represent sequences of the phonetic units, including selecting the optimum ones of such sequences to produce a selection signal representing recognition of the utterance,

said method being particularly characterized in that

the accessing step includes

assigning a phonetic unit signal and a phonetic duration signal from the stored model to one or more of said time frame portions of speech in response to one or more of said respective sets of acoustic feature signals, and

maximizing independently of the stored lexical candidates the likelihoods of each phonetic unit and each phonetic duration jointly with the likelihood of observing said one or more of said respective sets of acoustic feature signals,

said assigning and maximizing being performed recursively for all assignments and transitions over all time frames up to and including the present time frame; and

then retracing the actual maximization results by stepping through the phonetic determinations in a strict order to produce a proposed phonetic sequence for accessing the lexical candidates, and

subsequently accessing the stored lexical candidates with the proposed phonetic sequence to obtain signals representing a set of proposed lexical candidates, from which signals a final selection signal can be obtained.

2. A method for the recognition of speech, of the type claimed in claim 1,

said method being further characterized in that

the model storing step includes storing an ergodic model in which any state can occur after any other state, the model including examples of all such sequences and the corresponding transition probability signals.

3. A method for the recognition of speech, of the type claimed in claim 2;

said method being further characterized in that

the lexical candidate storing step comprises

storing words represented by a phonetic orthography which is characterized by partial phonetic information, such that words may be retrieved on the basis of the phonetic units which they contain, and

storing information linking a plurality of words containing like sequences of phonetic units, whereby for each sequence of phonetic units as many words as contain them are directly accessible.

4. A method for the recognition of speech, of the type including the steps of

storing signals representing a model of the language to be recognized, said model being of the state-transitional type, each state being uniquely identified with a phonetic unit,

each state having associated with it a portion of a transition matrix which describes which states can follow it and with what probability,

each state having associated with it an observational density function assigning to each set of speech feature signals that my be observed in fluent speech a likelihood of being observed in association with that state,

each state having associated with it a durational density function assigning to each duration it may have a likelihood of occurrence in fluent speech;

storing signals representing lexical candidates, said lexical candidates being assemblages of phonetic units of the language in association with partial phonetic information of the type found in dictionaries;

sequentially converting successive time frame portions of an utterance into signals representing respective sets of acoustic feature signals representative of the portions; and

accessing the stored model and stored lexical candidates to obtain signals which represent sequences of the phonetic units, including selecting the optimum ones of such sequences to produce a selection signal representing recognition of the utterance,

said method being particularly characterized in that

the accessing step includes

first accessing the stored model to obtain signals which represent proposed sequences of phonetic units independently of the stored lexical candidates, and

second accessing the stored lexical candidates in response to portions of the proposed sequences of phonetic units to obtain proposed lexical candidates each containing each said portion, including

whenever multiple proposed lexical candidates contain the same one said portion, aligning said multiple proposed lexical candidates with the one or more respective sets of acoustic feature signals from which said proposed sequences of phonetic units were obtained to evaluate said multiple proposed lexical candidates.

5. A method for the recognition of speech, of the type claimed in claim 4,

said method being further characterized in that

the model storing step includes storing an ergodic model in which any state can occur after any other state.

6. A method for the recognition of speech, of the type claimed in claim 5,

said method being further characterized in that

the aligning step evaluates said multiple proposed lexical candidates to include only those suitable for subsequent unambiguous ranking by processing by techniques that relate to sentence structure and meaning.

7. A method for the recognition of speech, of the type claimed in claim 5,

said method being further characterized in that

the aligning step evaluates said multiple proposed lexical candidates to select only the best one, whereby a selection signal representing the utterance as a word is produced.

8. Apparatus for the recognition of speech, of the type comprising

means for storing signals representing a model of the language to be recognized, said model being of the state-transitional type, each state being uniquely identified with a phonetic unit,

each state having associated with it a portion of a transition matrix which describes which states can follow it and with what probability,

each state having associated with it an observational density function assigning to each set of speech feature signals that my be observed in fluent speech a likelihood of being observed in association with that state,

each state having associated with it a durational density function assigning to each duration it may have a likelihood of occurrence in fluent speech;

means for storing signals representing lexical candidates, said lexical candidates being assemblages of phonetic units of the language in association with partial phonetic information of the type found in dictionaries;

means for sequentially converting successive time frame portions of an utterance into signals representing respective sets of acoustic feature signals representative of the portions; and

means for accessing the stored model and stored lexical candidates to obtain signals which represent sequences of the phonetic units, including means for selecting the optimum ones of such sequences to produce a selection signal representing recognition of the utterance,

said apparatus being particularly characterized in that

the accessing means includes

means for assigning a phonetic unit signal and a phonetic duration signal from the stored model to one or more of said time frame portions of speech in response to one or more of said respective sets of acoustic feature signals, and

means for maximizing independently of the stored lexical candidates the likelihoods of each phonetic unit and each phonetic duration jointly with the likelihood of observing said one or more of said respective sets of acoustic feature signals,

said assigning means and maximizing means being adapted to operate recursively for all assignments and transitions over all time frames up to and including the present time frame; and

said accessing means further includes

means for retracing the actual maximization results by stepping through the phonetic determinations in a strict order to produce a proposed phonetic sequence for accessing the lexical candidates, and

means for subsequently accessing the stored lexical candidates with the proposed phonetic sequence to obtain signals representing a set of proposed lexical candidates, from which signals a final selection signal can be obtained.

9. Apparatus for the recognition of speech, of the type claimed in claim 8,

said apparatus being further characterized in that

the means for storing a model includes means for storing an ergodic model in which any state can occur after any other state, the model including examples of all such sequences and the corresponding transition probability signals.

10. Apparatus for the recognition of speech, of the type claimed in claim 9;

said apparatus being further characterized in that

the lexical candidate storing means comprises

means for storing words represented by a phonetic orthography which is characterized by partial phonetic information, such that words may be retrieved on the basis of the phonetic units which they contain, and

means for storing information linking a plurality of words containing like sequences of phonetic units, whereby for each sequence of phonetic units as many words as contain them are directly accessible.

11. Apparatus for the recognition of speech, of the type comprising

means for storing signals representing a model of the language to be recognized, said model being of the state-transitional type, each state being uniquely identified with a phonetic unit,

each state having associated with it a portion of a transition matrix which describes which states can follow it and with what probability,

each state having associated with it an observational density function assigning to each set of speech feature signals that may be observed in fluent speech a likelihood of being observed in association with that state,

each state having associated with it a durational density function assigning to each duration it may have a likelihood of occurrence in fluent speech;

means for storing signals representing lexical candidates, said lexical candidates being assemblages of phonetic units of the language in association with partial phonetic information of the type found in dictionaries;

means for sequentially converting successive time frame portions of an utterance into signals representing respective sets of acoustic feature signals representative of the portions; and

means for accessing the stored model and stored lexical candidates to obtain signals which represent sequences of the phonetic units, including means for selecting the optimum ones of such sequences to produce a selection signal representing recognition of the utterance,

said apparatus being particularly characterized in that

the accessing means includes

means for first accessing the stored model to obtain signals which represent proposed sequences of phonetic units independently of the stored lexical candidates, and

means for next accessing the stored lexical candidates in response to portions of the proposed sequences of phonetic units to obtain proposed lexical candidates each containing each said portion, and

the accessing means further includes

means for aligning said proposed lexical candidates, whenever more than one exists, with the one or more respective sets of acoustic feature signals from which said proposed sequences of phonetic units were obtained to reduce the number of said proposed lexical candidates.

12. Apparatus for the recognition of speech, of the type claimed in claim 11,

said apparatus being further characterized in that

the model storing means includes means for storing an ergodic model in which any state can occur after any other state.

13. Apparatus for the recognition of speech, of the type claimed in claim 12,

said apparatus being further characterized in that

the aligning means evaluates said multiple proposed lexical candidates to include only those suitable for subsequent unambiguous ranking by techniques that relate to sentence structure and meaning.

14. Apparatus for the recognition of speech, of the type claimed in claim 12,

said apparatus being further characterized in that

the aligning means evaluates said multiple proposed lexical candidates to select only the best one, whereby the aligning means is capable of producing a selection signal representing the utterance as a word.