15th Annual Conference of the International Speech Communication Association

September 14-18, 2014

Regularized Feature-Space Discriminative Adaptation for Robust ASR

Takashi Fukuda (1), Osamu Ichikawa (1), Masafumi Nishimura (1), Steven J. Rennie (2), Vaibhava Goel (2)

(1) IBM Research Tokyo, Japan
(2) IBM T.J. Watson Research Center, USA

Model-space adaptation techniques such as MLLR and MAP are often used for porting old acoustic models into new domains. Discriminative schemes for model adaptation based on MMI and MPE objective functions are also utilized. For feature-space adaptations, one extension to the well-known feature-space discriminative training (fMPE) algorithm, feature-space discriminative adaptation, was recently proposed to adapt fMPE transforms. Feature-space discriminative adaptation was shown to work well for some situations when sufficient adaptation data is available. This paper improves the feature-space discriminative adaptation by introducing a regularization term for an indirect differential computation of the fMPE objective function, and also by updating the acoustic models with MAP instead of ML criterion during the adaptation. The proposed method performed favorably for the adaptation conditions from general-purpose LVCSR to automotive environments with small amounts of adaptation data, and yielded 4.4% relative improvement as compared with MAP-adapted system without using the fMPE adaptation.

Full Paper

Bibliographic reference.  Fukuda, Takashi / Ichikawa, Osamu / Nishimura, Masafumi / Rennie, Steven J. / Goel, Vaibhava (2014): "Regularized feature-space discriminative adaptation for robust ASR", In INTERSPEECH-2014, 2185-2188.