TOWARDS SPEAKER AND ENVIRONMENTAL ROBUSTNESS IN ASR: THE HIWIRE PROJECT

In this paper, we present algorithms for dealing with variability and mismatch in speech recognition due to environmental conditions and non-native speaker populations. The proposed algorithms cover a broad spectrum of ideas including robust feature extraction, fea-ture compensation and speech enhancement. Specifically the fol-lowing algorithms are presented and evaluated: beamforming for multi-microphone speech recognition, robust modulation and frac-tal features, Teager energy cepstrum coefficients, parametric feature equalization, speech enhancement, and acoustic modeling for non-native speech recognition. Also the problem of feature fusion and voice activity detection are discussed. Evaluation results on the AU-RORA databases under the auspices of the HIWIRE project show that significant gains can be achieved under adverse or mismatched conditions using these algorithms. Relative error rate reduction of up to 50 % was shown for multi-microphone ASR, robust feature com-bination and speech enhancement. 30-40 % reduction was shown for parametric feature equalization and non-native acoustic models. 1