Tetsuya Takiguchi, Qiang Huo, Satoshi Nakamura, Kiyohiro Shikano
Adaptation of Model Parameters by HMM
Decomposition in Reverberant Environments
Abstract:An HMM composition method has been reported in [5] to cope with the distant-talking speech recognition problem in noisy and/or reverberant environments. In
order to estimate the HMM parameters of the acoustic transfer function of the
reverberant room, actual impulse responses at different positions of the room have
to be measured first. It turns out inconvenient and unrealistic to measure the
impulse responses for every possible testing room. In this report, we study a method
for estimating the HMM parameters of the acoustic transfer function by using
a small amount of adaptation data (possibly derived from actual testing data).
The proposed technique is based on HMM decomposition which can be viewed
as an inverse process of the HMM composition. We will report some preliminary
experimental results to show how it works in a simulated distant-talking speech
recognition scenario using a single omni-directional microphone.