Eigenvoice re-estimation technique of acoustic models for speech recognition, speaker identification and speaker verification
View Patent ↗A reduced dimensionality eigenvoice analytical technique is used during training to develop context-dependent acoustic models for allophones. Re-estimation processes are performed to more strongly separate speaker-dependent and speaker-independent components of the speech model. The eigenvoice technique is also used during run time upon the speech of a new speaker. The technique removes individual speaker idiosyncrasies, to produce more universally applicable and robust allophone models. In one embodiment the eigenvoice technique is used to identify the centroid of each speaker, which may then be “subtracted out” of the recognition equation.
1. A method for developing context dependent acoustic models, comprising the steps of:
developing a low-dimensional space from training speech data obtained from a plurality of training speakers by constructing an eigenspace from said training speech data;
representing the training speech data from each of said plurality of training speakers as the combination of a speaker dependent component and a speaker independent component;
representing said speaker dependent component as centroids within said low-dimensional space;
representing said speaker independent component as linear transformations of said centroids; and
performing maximum likelihood re-estimation on said training speech data of at least one of said low-dimensional space, said centroids, and said linear transformations to represent context dependent acoustic model.
2. The method of claim 1 wherein said training speech data is separated by identifying context dependent data and using said context dependent data to identify said speaker independent data.
3. The method of claim 1 wherein said training speech data is separated by identifying context independent data and using said context independent data to identify said speaker dependent data.
4. The method of claim 1 wherein said maximum likelihood re estimation step is performed iteratively.
5. The method of claim 1 wherein said linear transformations are effected as an offsets from said centroids.
6. The method of claim 1 wherein said maximum likelihood re-estimation step generates a re-estimated low-dimensional space, re-estimated centroids and re-estimated offsets and wherein said context dependent acoustic models are constructed using said re-estimated low-dimensional space and said re-estimated offsets.
7. The method of claim 1 wherein said linear transformations of said centroids are represented in tree data structures corresponding to individual sound units.
8. The method of claim 5 wherein said offsets are represented in tree data structures corresponding to individual sound units.
9. The method of claim 1 further comprising:
using said speaker dependent component to perform speaker verification.
10. The method of claim 1 further comprising:
using said speaker dependent component to perform speaker identification.
11. A method for developing context dependent acoustic models, comprising the steps of:
developing a low-dimensional space from training speech data obtained from a plurality of training speakers by constructing an eigenspace from said training speech data;
representing the training speech data from each of said plurality of training speakers as the combination of a speaker dependent component and a speaker independent component;
representing said speaker dependent component as centroids within said low-dimensional space;
representing said speaker independent component as linear transformations of said centroids; and
performing maximum likelihood re-estimation on said training speech data of at least one of said low-dimensional space, said centroids, and said linear transformations to represent context dependent acoustic model, wherein said linear transformations are effected as offsets from said centroids, said maximum likelihood re-estimation step generates a re-estimated low-dimensional space, re-estimated centroids and re-estimated offsets and wherein said context dependent acoustic mociels are constructed using said re-estimated low-dimensional space and said re-estimated offsets.
12. The method of claim 11 wherein said linear transformations of said centroids are represented in tree data structures corresponding to individual sound units.
13. The method of claim 11 wherein said offsets are represented in tree data structures corresponding to individual sound units.
14. The method of claim 11 further comprising:
using said speaker dependent component to perform speaker verification.
15. The method of claim 11 further comprising:
using said speaker dependent component to perform speaker identification.
16. The method of claim 11 wherein said training speech data is separated by identifying context dependent data and using said context dependent data to identify said speaker independent data.
17. The method of claim 11 wherein said training speech data is separated by identifying context independent data and using said context independent data to identify said speaker dependent data.
18. The method of claim 11 wherein said maximum likelihood re-estimation step is performed iteratively.