EUROSPEECH 2001 Scandinavia
The present paper reports on a novel technique for the identification and replacement of spectral coefficients degraded by impulsive noise. The problem is viewed from the perspective of Missing Feature Theory (MFT). The analysis is carried out in the linear spectrum prior to, or after applying the mel-scale filter-bank depending on whether one aims at improving the quality of perception of speech recordings or at Automatic Speech Recognition (ASR). Each filter-bank output is considered to be a time series drawn from an Auto-Regressive process (AR). A validation corpus of undistorted recordings is used to derive a-priori bounds on the expected prediction error of each AR model. In operational conditions, the prediction procedure is monitored and the violation of the statistical bounds indicates band corruption and entails the substitution of the degraded spectral coefficients by the prediction of the corresponding AR model. ASR experiments and informal listening tests demonstrate large improvement in terms of word recognition performance and Itakura-Saito divergence at very low Signal to Impulsive Noise Ratios (SINRs). Data, and implementation code can be found at: ftp://wcl.ee.upatras.gr/
Bibliographic reference. Potamitis, I. / Fakotakis, Nikos (2001): "Autoregressive time-frequency interpolation in the context of missing data theory for impulsive noise compensation", In EUROSPEECH-2001, 2475-2478.