EUROSPEECH 2003 - INTERSPEECH 2003
In this paper, two low-complexity histogram equalization algorithms are presented that significantly reduce the mismatch between training and testing conditions in HMM-based automatic speech recognizers. The proposed algorithms use Gaussian approximations for the initial and target distributions and perform a linear mapping between them. We show that even this simplified mapping can improve the noise robustness of ASR systems, while the associated computational load, memory requirements, and algorithmic delay are minimal. The proposed algorithms were evaluated in a multi-lingual speaker independent isolated word recognition task without and in combination with on-line MAP acoustic model adaptation. The best results obtained showed an approximate 25/20% relative error-rate reduction without/with acoustic model adaptation.
Bibliographic reference. Haverinen, Hemmo / Kiss, Imre (2003): "On-line parametric histogram equalization techniques for noise robust embedded speech recognition", In EUROSPEECH-2003, 3061-3064.