This thesis consists of three parts. We first evaluate automatic speech recognition (ASR) results at various locations of a voice communication system, and determine causes for performance deterioration compared to traditional ASR where the original speech signal is available. We then proceed to alleviate two feature distortions responsible for the performance degradation. The first comes from the fact that all model-based speech coding standards use linear prediction (LP) parameters to represent spectral information, whereas the Mel-frequency cepstral coefficients (MFCCs) are generally more robust for ASR purpose. The second distortion is due to complete loss of information packets. Since such loss tends to happen in bursts due to fading channels in wireless communication or network congestion in voice over IP, most frame-concealment techniques employed by the decoder, including insertion, interpolation, or substitution, are futile in trying to smooth out the gap. The ultimate goal is to make appropriate modification at the transmitter/encoder side, and provide the most suitable and reliable feature sets to restore the performance of a given ASR algorithm at the receiver/client side to near the level when original data is accessible. Since it is difficult to imagine any voice service excluding subjective listening, the modifications should lead to minimal deterioration in perceptual quality.
展开▼
