ディミトリ・リティシェフ,シャン・ホウ, ハラルド・シンガー
Implementation and Testing of
Quasi-Bayes Speaker Adaptation
Algorithm
Abstract:The Quasi-Bayes speaker adaptation algorithm enables incremental adaptation of HMM parameters as speech from the target speaker becomes available, without requiring that the adaptation speech be stored for subsequent adaptation. The algorithm has been implemented as part of
the ATR SSS-ToolKit recognition system and evaluated on the ATR spontaneous speech database.
Experimental results indicate that the Quasi-Bayes adaptation algorithm is a computationally efficient method for achieving consistent improvement in recognition accuracy in supervised mode,
i.e., relying on manually prepared transcriptions of the adaptation speech. The results also show
that the algorithm does not yield consistent improvement in recognition accuracy in unsupervised
mode. Additional research is necessary to achieve unsupervised adaptation capability.