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