Speech coding and joint data/channel bias estimation using finite state vector quantizer derived from sequential constraints

Claim 1

A method for conditioning a speech signal to compensate said signal for unknown distortion components included therein, wherein said conditioning is related to a coding of said signal based on sequential constraints characterizing said signal, said method comprising the steps of: selecting a set of models for general acoustic modeling of human speech; decomposing said model set into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; operating on said speech feature vector centroids in each said decomposed acoustic class with a quantizing means to determine at least one code word corresponding to each said decomposed acoustic class; constructing a finite state vector quantizer based on said code words so determined; using said finite state vector quantizer to decode speech feature vectors of said signal, thereby determining which portions of said signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; estimating a bias term in said signal along said minimum sequence distortion path; and subtracting said estimated bias term from said signal.

Claim 2

A method for conditioning a speech signal to compensate said signal for unknown distortion components included therein, wherein said conditioning is related to a coding of said signal based on sequential constraints characterizing said signal, said method comprising the steps of: selecting a set of models for general acoustic modeling of human speech; decomposing said model set into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; operating on said speech feature vector centroids in each said decomposed acoustic class with a quantizing means to determine at least one code word corresponding to each said decomposed acoustic class; constructing a finite state vector quantizer based on said code words so determined; using said finite state vector quantizer to decode speech feature vectors of said signal, thereby determining which portions of said signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; estimating a bias term in said signal along said minimum sequence distortion path; and subtracting said estimated bias term from said signal.

Claim 10

A method for providing a coding for a speech signal, wherein said coding embeds sequential constraints characterizing human speech comprising the steps of: selecting a set of models for general acoustic modeling of human speech; decomposing said model set into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; and operating on said speech feature vector centroids in each said decomposed acoustic class with a quantizing means to determine at least one code word corresponding to each said decomposed acoustic class.

Claim 11

A method for providing a coding for a speech signal, wherein said coding embeds sequential constraints characterizing human speech comprising the steps of: selecting a set of models for general acoustic modeling of human speech; decomposing said model set into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; and operating on said speech feature vector centroids in each said decomposed acoustic class with a quantizing means to determine at least one code word corresponding to each said decomposed acoustic class.

Claim 17

A method for reducing the effect of an unknown signal bias in an input speech signal, wherein a set of code words which embed sequential constraints inherent in said speech signal is generated, and further wherein said code word set is incorporated in a finite state vector quantizer as a code constellation thereof, said method comprising the steps of: using said finite state vector quantizer to decode speech feature vectors of said speech signal, thereby determining which portions of said speech signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; estimating a bias term in said speech signal along said minimum sequence distortion path; and subtracting said estimated bias term from said speech signal.

Claim 18

A method for reducing the effect of an unknown signal bias in an input speech signal, wherein a set of code words which embed sequential constraints inherent in said speech signal is generated, and further wherein said code word set is incorporated in a finite state vector quantizer as a code constellation thereof, said method comprising the steps of: using said finite state vector quantizer to decode speech feature vectors of said speech signal, thereby determining which portions of said speech signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; estimating a bias term in said speech signal along said minimum sequence distortion path; and subtracting said estimated bias term from said speech signal.

Claim 24

A system for reducing the effect of an unknown signal bias in an input speech signal, wherein said system operates on a coding of said signal based on sequential constraints characterizing said signal, said system comprising: means for decomposing a selected set of models describing acoustic events of human speech into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; a first processing means for determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; a second processing means operating on said speech feature vector centroids in each said decomposed acoustic class to implement a quantizing process to thereby determine at least one code word corresponding to each said decomposed acoustic class; a means for constructing a finite state vector quantizer based on said code words so determined; a third processing means for causing said finite state vector quantizer to decode speech feature vectors of said signal, thereby determining which portions of said signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; a fourth processing means for estimating a bias term in said signal along said minimum sequence distortion path; and a fifth processing means for subtracting said estimated bias term from said signal; and wherein one or more of said first through said fifth processing means may be combined into a common processing means.

Claim 25

A system for reducing the effect of an unknown signal bias in an input speech signal, wherein said system operates on a coding of said signal based on sequential constraints characterizing said signal, said system comprising: means for decomposing a selected set of models describing acoustic events of human speech into a plurality of acoustic classes, said classes reflecting a general correspondence with said sequential constraints; a first processing means for determining a set of centroids for clusters of training speech feature vectors included in each said decomposed acoustic class; a second processing means operating on said speech feature vector centroids in each said decomposed acoustic class to implement a quantizing process to thereby determine at least one code word corresponding to each said decomposed acoustic class; a means for constructing a finite state vector quantizer based on said code words so determined; a third processing means for causing said finite state vector quantizer to decode speech feature vectors of said signal, thereby determining which portions of said signal correspond to each of said code words, said decoded speech feature vectors collectively comprising a minimum sequence distortion path in said finite state vector quantizer; a fourth processing means for estimating a bias term in said signal along said minimum sequence distortion path; and a fifth processing means for subtracting said estimated bias term from said signal; and wherein one or more of said first through said fifth processing means may be combined into a common processing means.