Bell Telephone Laboratories

At&t Bell Laboratories

American Telephone & Telegraph Co., At&t Bell Laboratories

N.v. Philips' Gloeilampenfabrieken

Stephen E Levinson

Lawrence Richard Rabiner

Man M Sondhi

Biing-Hwang Juang

Bernd E Aldefeld

Frank C Pirz

Allen L Gorin

Thomas G Szymanski

A machine-based learning system receives inputs and proposes outputs using a network of connection weights to relate the inputs to the outputs. When a user affirms a proposed output, the system adjusts the connection weights to strengthen the relationship between elements of the inputs and that output. Learning occurs in the course of successive iterations as the connection weights adapt to relationships between semantically-significant input elements and related outputs. The system can be used to acquire language from inputs provided in the form of text or speech such that the connection weights adapt to semantically-significant words without need for defining word meanings. Such language acquisition contrasts with previous systems in which words are predefined.

Machine-based learning system

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.

Speech recognition by acoustic/phonetic system and technique

Markov model speech pattern templates are formed for speech analysis systems by analyzing identified speech patterns to generate frame sequences of acoustic feature signals representative thereof. The speech pattern template is produced by iteratively generating succeeding Markov model signal sets starting with an initial Markov model signal set. Each iteration includes forming a set of signals representative of the current iteration Markov model of the identified speech pattern responsive to said frame sequences of acoustic feature signals and one of the previous Markov model signal sets and comparing the current iteration Markov model signal set with said previous Markov model signal set to generate a signal corresponding to the similarity therebetween. The iterations are terminated when said similarity signal is equal to or smaller than a predetermined value and the last formed Markov model signal set is selected as a reference template for said identified speech pattern. The state transition model has increased accuracy by grouping the feature signals into related clusters corresponding to states of the previous state transitional model, whereby with further grouping of the feature signals the continuous probability density function acquires components representing a mixture of different continuous probability density functions.

Hidden Markov model speech recognition arrangement

A speech recognizer includes a plurality of stored constrained hidden Markov model reference templates and a set of stored signals representative of prescribed acoustic features of the said plurality of reference patterns. The Markov model template includes a set of N state signals. The number of states is preselected to be independent of the reference pattern acoustic features and preferably substantially smaller than the number of acoustic feature frames of the reference patterns. An input utterance is analyzed to form a sequence of said prescribed feature signals representative of the utterance. The utterance representative prescribed feature signal sequence is combined with the N state constrained hidden Markov model template signals to form a signal representative of the probability of the utterance being each reference pattern. The input speech pattern is identified as one of the reference patterns responsive to the probability representative signals.

A string of input characters is recognized as one of a plurality of reference sets which are stored in predetermined groups. A set of signals representative of the correspondence of the input characters to the reference characters is generated. The input character string is identified responsive to the character correspondence signals. The character string identification comprises forming a set of signals each representative of the correspondence of the input character string to one of the predetermined groups responsive to the character correspondence signals. Jointly responsive to the character correspondence signals, the reference character sets and the group correspondence signals, the reference set that most closely corresponds to the input character string is selected.

Spelled word recognizer

The accuracy of segmenting an utterance into words is improved by the use of sequence-defining signals. An utterance is recognized as one of a plurality of reference word sequences in an arrangement wherein a set of signals is generated to define the syntax, i.e., word arrangements, of the sequences. Each sequence corresponds to a selected series of reference words. A signal is generated to identify the successive word positions of the sequences. Responsive to the sequence defining signals, the utterance, and the reference words, a set of signals is produced, each representing the correspondence between said utterance and one of the sequences. The sequence having the closest correspondence to the utterance is selected. The sequence correspondence signal generation includes selecting, in each identified word position, the word position reference word for each sequence and the portion of the utterance corresponding thereto responsive to said sequence defining signals, and generating a signal representative of the acoustic correspondence between each sequence word position reference word and its selected utterance portion.

Syntactic continuous speech recognizer

A syntactic analyzer recognizes a series of spoken words as one of a set of predetermined sequences of prescribed words defined by a state sequence array of state transition signals representing state linked prescribed words. Each sequence corresponds to a selected plurality of linked state transition signals ending in a final state. For each word position of the series, a set of signals representative of the acoustic correspondence between the position word and the array prescribed words is generated. A cumulative correspondence signal is produced from the array state transition signals and the correspondence signals. Upon termination of the last word position, the sequence having the closest cumulative correspondence to the spoken word series is identified.

Growing community of inventors

Westfield, NJ, United States of America

Stephen E Levinson