TR2014-022

Deep recurrent de-noising auto-encoder and blind de-reverberation for reverberated speech recognition

    •  Weninger, F., Watanabe, S., Tachioka, Y., Schuller, B., "Deep Recurrent De-noising Auto-encoder and Blind De-reverberation for Reverberated Speech Recognition", IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), DOI: 10.1109/​ICASSP.2014.6854478, May 2014, pp. 4623-4627.
      BibTeX TR2014-022 PDF
      • @inproceedings{Weninger2014may1,
      • author = {Weninger, F. and Watanabe, S. and Tachioka, Y. and Schuller, B.},
      • title = {Deep Recurrent De-noising Auto-encoder and Blind De-reverberation for Reverberated Speech Recognition},
      • booktitle = {IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
      • year = 2014,
      • pages = {4623--4627},
      • month = may,
      • publisher = {IEEE},
      • doi = {10.1109/ICASSP.2014.6854478},
      • url = {https://www.merl.com/publications/TR2014-022}
      • }
  • Research Areas:

    Artificial Intelligence, Speech & Audio

TR Image
Flowchart of the proposed method. Dashed lines depict optional processing steps. FB: filterbank. (*) Linear transformations: DCT (to obtain MFCC), LDA, MLLT, CMLLR – see text.
Abstract:

This paper describes our joint efforts to provide robust automatic speech recognition (ASR) for reverberated environments, such as in hands-free human-machine interaction. We investigate blind feature space de-reverberation and deep recurrent de-noising auto-encoders (DAE) in an early fusion scheme. Results on the 2014 REVERB Challenge development set indicate that the DAE front-end provides complementary performance gains to multi-condition training, feature transformations, and model adaptation. The proposed ASR system achieves word error rates of 17.62 % and 36.6 % on simulated and real data, which is a significant improvement over the Challenge baseline (25.16 and 47.2 %).