Analysis of different acoustic front-ends for automatic voice over IP recognition

We investigated the usage for automatic speech recognition of different acoustic features, obtained from the output bitstream of a voice over IP codec. In particular, we analyzed the influence, on recognition performance, of both analysis rate and vector quantization of acoustic parameters introduced by the codec. Particular care has to be taken to train acoustic models at the reduced analysis rate employed by the codec: some related issues are discussed in the paper. We also used a model for simulating packet loss and we measured the corresponding performance degradation. Experiments were carried out on both clean and noisy speech databases.