Acoustic Model Adaptation on Speech and Audio Coding Distortion

This paper studies the influence of coding distortion on speech recognition performance. The effects of the coding-decoding process on the Gaussian distributions were investigated by statistics comparison of clean and coded speech data. Further, several different coders were used for speech recognition performance assessment of the UASR (Unified Automatic Speech Recognition and Synthesis) system. The assessment was performed firstly on acoustic model trained on clean speech and latter on coder adapted acoustic models.