Noise-robust cellular phone speech recognition using codec-adapted speech and noise models

In this paper, we introduce a rapid and noise-robust CODEC adaptation for cellular phone speech recognition in adverse environment. Our proposal is to use CODEC-dependent noise models (CDNM) in addition to CODEC-dependent speech models (CDSM) for reducing speech detection errors caused by the background noise and compensating nonlinear speech distortion due to the low-bitrate CODECs. In this approach, the CODEC type is automatically determined through the recognition process by comparing likelihoods of CODEC-dependent models. Experiments on 3k words recognition of noisy speech showed 25% improvement in accuracy, and 46% reduction in early starting point detection errors over the conventional CODEC-independent approach.