Inference of Missing or Degraded Data for Noise Robust Speech Processing
Bengt Borgström · 2010
OF THE DISSERTATION Inference of Missing or Degraded Data for Noise Robust Speech Processing by Bengt Jonas Borgstrom Doctor of Philosophy in Electrical Engineering University of California, Los Angeles, 2010 Professor Abeer Alwan, Chair In real world speech processing systems, speech signals are often corrupted by background acoustic noise or reverberation. Additionally, for systems which involve transmission of speech data over error-prone communication channels, signals may suffer from packet loss. This dissertation addresses two general frameworks for which compensation of corruptive acoustic noise and channel errors can benefit performance, namely remote speech communication and automatic speech recognition. In the case of ASR, front-end missing feature (MF) spectral reconstruction is explored. Two solutions are offered, the first of which uses HMM-based processing and accounts for temporal and/or frequency correlation. The second exploits the sparsity of spectrographic speech data to formulate the reconstruction problem as a linear program. Each approach is successfully applied in both the Mel-filtered and log Mel-filtered domains. Finally, a statistical approach to Mel-domain mask estimation is proposed, which is used to differentiate between reliable and unreliable time-frequency components. Theory de-