Edward Willems, Tetsuo Kosaka, Jun-Ichi Takami,
Shigeki Sagayama
A Study on Dynamic Speaker
Adaptation using HMM-Nets
Abstract:This report describes a new approach to dynamic speaker adaptation, which relies on
switching between different methods of adaptation in order to gain maximum performance
depending on the amount of speech data obtained through the speech recognition session.
The speech recognition performance of speaker adaptive systems is determined by the
specific method used for the adaptation as well as by the amount of available training speech data.
Furthermore, the effectiveness of the adaptation often depends on the speakers.
We present a two dynamic features to include in the design of speaker-adaptive recognisers
using Hidden Markov Networks : dynamic method selection and dynamic method adaptation.
We also present an implementation of the former feature : a system which can switch between
three different speaker adaptation techniques, namely, vector field smoothing, speaker-tied
weight training and speaker-free weight training. The methods are selected according to the
most likely candidate they produce, based on the input speech data. Experimental results
show that this dynamic system achieved better results compared to conventional recognisers
which use a single adaptation procedure.