Enhancing acoustic models for robust speaker verification

Acoustic model enhancement (AME) refers to adapting the acoustic models to compensate for the distortion induced by a speech enhancement technique. This work extends the AME technique for speaker verification recently presented by incorporating the corresponding adaptation of the model variances, and by exploring the trade off between noise over-estimation and flooring distortion in the verification error. By using spectral subtraction (SS) as the speech enhancement technique, the extended AME highly outperformed SS alone particularly at moderately low SNRs (0 dB - 15 dB), where the adaptation of the variance was found to considerably improve the equal error rate (EER).