ISCA Archive Interspeech 2006
ISCA Archive Interspeech 2006

Interleaving and MMSE estimation with VQ replicas for distributed speech recognition over lossy packet networks

Angel M. Gómez, Antonio M. Peinado, Victoria Sánchez, José L. Carmona, Antonio J. Rubio

In this work we evaluate the performance of MMSE estimation with a media-specific FEC based on VQ replicas in comparison with MAP estimation and interleaving, both operating in a DSR system over a loss-prone packet switched network. Both schemes combine a sender-driven with a receiver-based technique and, as we show, clearly outperform the standard Aurora mitigation. However, as independent techniques, interleaving and FEC codes could be jointly applied. Although this would provide better results, a direct combination of FECs and interleaving involves a sum of the delays of both operations. In this work, we introduce a double stream-based strategy that avoids this sum of delays.