Method and apparatus of hierarchically organizing an acoustic model for speech recognition and adaptation of the model to unseen domains
AssigneeLernout & Hauspie Speech Products N.V.
InventorsAlex Waibel, Juergen Fritsch
Life of this patent
Estimating the term…
Abstract
A method of organizing an acoustic model for speech recognition is comprised of the steps of calculating a measure of acoustic dissimilarity of subphonetic units. A clustering technique is recursively applied to the subphonetic units based on the calculated measure of acoustic dissimilarity to automatically generate a hierarchically arranged model. Each application of the clustering technique produces another level of the hierarchy with the levels progressing from the least specific to the most specific. A technique for adapting the structure and size of a trained acoustic model to an unseen domain using only a small amount of adaptation data is also disclosed.
Claims36
A memory device encoded with a computer program for enabling a computer program to execute a method comprised of the steps of:
- initializing a system of clusters, each cluster containing a model of the class-conditional likelihood and a count indicating the frequency of that class in the training set;
- computing within cluster priors for each cluster using the counts;
- computing a divergence measure between all pairs of clusters;
- determining the pair of clusters with a minimum divergence measure;
- creating a new cluster based on the determined pair and deleting the pair having a minimum divergence measure;
- and
- repeating the process until some predetermined criteria is met.
An apparatus, comprising:
- means for initializing a system of clusters, each cluster containing a model of the class-conditional likelihood and a count indicating the frequency of that class in the training set;
- means for computing within cluster priors for each cluster using the counts;
- means for computing a divergence measure between all pairs of clusters;
- means for determining the pair of clusters with a minimum divergence measure;
- and
- means for creating a new cluster based on the determined pair and deleting the pair having a minimum divergence measure.
A method of organizing an acoustic model for speech recognition, comprising:
- initializing a system of clusters, each cluster containing a model of the class-conditional likelihood and a count indicating the frequency of that class in the training set;
- computing within cluster priors for each cluster using the counts;
- computing a divergence measure between all pairs of clusters;
- determining the pair of clusters with a minimum divergence measure;
- creating a new cluster based on the determined pair and deleting the pair having a minimum divergence measure;
- and
- repeating the process until some predetermined criteria is met.
A method of structurally adapting a hierarchical acoustic model having nodes and leaves to a new domain, comprising:
- identifying nodes that receive more than a predetermined amount of adaptation data;
- adapting the local estimators of conditional posteriors and priors of the identified nodes using data from the new domain;
- removing a predetermined number of the non-identified nodes;
- creating new leaves where needed to replace the removed nodes;
- and
- relating all HMM states corresponding to the new leaves such that they share a single model represented by the new leaves.
A memory device encoded with a computer program for enabling a computer program to execute a method for structurally adapting a hierarchical acoustic model having nodes and leaves to a new domain comprised of the steps of:
- identifying nodes that receive more than a predetermined amount of adaptation data;
- adapting the local estimators of conditional posteriors and priors of the identified nodes using data from the new domain;
- removing a predetermined number of the non-identified nodes;
- creating new leaves where needed to replace the removed nodes;
- and
- relating all HMM states corresponding to the new leaves such that they share a single model represented by the new leaves.
An apparatus for structurally adapting a hierarchical acoustic model having nodes and leaves to a new domain, comprising:
- means for identifying nodes that receive more than a predetermined amount of adaptation data;
- means for adapting the local estimators of conditional posteriors and priors of the identified nodes using data from the new domain;
- means for removing a predetermined number of the non-identified nodes;
- means for creating new leaves where needed to replace the removed nodes;
- and
- means for relating all HMM states corresponding to the new leaves such that they share a single model represented by the new leaves.
Dependencies are inferred from the claim text.