9th Annual Conference of the International Speech Communication Association

Brisbane, Australia
September 22-26, 2008

Combining Noise Compensation and Missing-Feature Decoding for Large Vocabulary Speech Recognition in Noise

Jianhua Lu, Ji Ming, Roger Woods

Queen's University Belfast, UK

In this paper we propose a combination of noise compensation and missing-feature decoding for large-vocabulary speech recognition in noisy environments. Two approaches for noise compensation have been studied. These are noise training and vector Taylor series expansion, aiming to compensate white Gaussian noise at various levels. This is followed by subband missing-feature decoding to reduce the model/data mismatch. The proposed approach requires little knowledge about the noisy environments. The Aurora 4 corpus is used for the experiments. Better results are obtained by the new approach over a multi-condition baseline system.

Full Paper

Bibliographic reference.  Lu, Jianhua / Ming, Ji / Woods, Roger (2008): "Combining noise compensation and missing-feature decoding for large vocabulary speech recognition in noise", In INTERSPEECH-2008, 1269-1272.