Speaker adaption of CDHMMs using Bayesian learning
The aim of the paper is to investigate and evaluate the on-line use of the Bayesian learning approach, known also as Maximum A Posteriori (MAP), to the speaker adaptation of Continuous Density Hidden Markov Models (CDHMMs). The parameters of the Gaussian mixture output densities are adapted using the exponential forgetting mechanism and performing the a priori estimation in a model-based outline. The model-based approach permits an easy hyper-parameter estimation and updating. The comparison between maximum likelihood and MAP with forgetting mechanism shows that Bayesian models outperform the models based on speaker-independent plus speaker-dependent maximum likelihood. Moreover, the MAP models are better than speaker-dependent models when the training data are below 20 minutes of speech. The cepstral mean normalization technique is used to compensate the variation of the transmission channel. For the entire collection see [Zbl 0927.00044].
- Transformation-based Bayesian prediction for adaptation of HMMs
- On-line incremental speaker adaptation for broadcast news transcription
- An on-line speaker adaptation method for HMM-based speech recognizers
- Recognizing 100 speakers using homologous naive Bayes
- Incremental MLLR speaker adaptation by fuzzy logic control
- A mixture of recurrent neural networks for speaker normalisation
- Incremental MLLR speaker adaptation by fuzzy logic control
- Adaptive hierarchy of hidden Markov models for transformation-based adaptation
- An on-line speaker adaptation method for HMM-based speech recognizers
- Recognizing 100 speakers using homologous naive Bayes
- On-line incremental speaker adaptation for broadcast news transcription
- Transformation-based Bayesian prediction for adaptation of HMMs
This page was built for publication: Speaker adaption of CDHMMs using Bayesian learning
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4942719)
