EUROSPEECH 2003 - INTERSPEECH 2003
Double-talk recognition under a distant microphone condition, a serious problem in speech applications in a real environment, is realized through use of modified SAFIA and acoustic model adaptation or training. The original SAFIA is a high-performance audio segregation method based on band selection using two directivity microphones. We have modified SAFIA by adopting array signal processing and have realized optimal directivity for SAFIA. We also used generalized harmonic analysis (GHA) instead of FFT for the spectral analysis in SAFIA to remove the effect of windowing which causes sound-quality degradation in SAFIA.
These modifications of SAFIA enable good segregation in a human auditory sense, but the quality is still insufficient for recognition. Because SAFIA causes some particular distortion, we used MLLR-based acoustic model adaptation and immunity training to be robust to the distortion of SAFIA. These efforts enabled 76.2% word accuracy under the condition that the SN ratio is 0 dB, this represents a 45% reduction in the error obtained in the case where only array signal processing was used, and a 30% error reduction compared with when only SAFIA-based audio segregation was used.
Bibliographic reference. Sekiya, Toshiyuki / Ogawa, Tetsuji / Kobayashi, Tetsunori (2003): "Speech recognition of double talk using SAFIA-based audio segregation", In EUROSPEECH-2003, 1285-1288.