- Abstract:
- This work introduces two methods for adapting the observation process parameters of linear dynamic models (LDM), a form of state-space model. The application which motivates this work is automatic speech recognition, though these techniques are applicable to many other domains in which linear-Gaussian models are used. We describe two approaches: the first uses the expectation-maximization (EM) algorithm to estimate transforms for location and covariance parameters, and the second uses a generalized EM approach which reduces computation in making updates from O(p^6) to O(p^3), where p is the feature dimension. We present the results of speaker adaptation on TIMIT phone classification and recognition experiments, and demonstrate relative error reductions of up to 6%. We find minimal differences in the results using the EM and GEM methods and therefore propose that the GEM approach might also prove useful for rapid adaptation of hidden Markov model (HMM) systems which use non-diagonal covariances.
- Links To Paper
- 1st Link
- Bibtex format
- @Article{EDI-INF-RR-1009,
- author = {
Jolyon Frankel
and Simon King
},
- title = {Observation process adaptation for linear dynamic models},
- journal = {Speech Communication},
- year = 2006,
- month = {Sep},
- volume = {48(9)},
- pages = {1192-1199},
- doi = {10.1016/j.specom.2006.05.001},
- url = {http://www.cstr.ed.ac.uk/downloads/publications/2006/Frankel_King_SPECOM2006.pdf},
- }
|
