Combining Bayesian learning and vector field smoothing for on-line incremental speaker adaptation

In this paper we investigate the combination of Bayesian Learning (also known as Maximum A Posteriori-MAP) and Vector Field Smoothing (VFS) to the on-line incremental speaker adaptation of Continuous Density Hidden Markov Models (CDHMMs). The parameters of the Gaussian mixture output densities are adapted during the MAP step, using the exponential forgetting mechanism and performing the a-priori parameter estimation in a model based outline. The unadapted MAP models are then reestimated using the VFS technique. Several tests compare the error rate reduction as a function of the incremental adaptation step and of the size of adaptation data for each step. The experiments were run on a speaker dependent continuous speech recognition task, for the Italian language, with a test vocabulary size of 247 words, without any language models. The initial speaker independent models were trained on PSTN speech data, while the adaptation data were collected in a quiet environment through a PBX connection. The cepstral mean normalization (CMN) was used to deal with the acoustic mismatch.