Matching the acoustic model to front-end signal processing for ASR in noisy and reverberant environments

Distant-talking automatic speech recognition (ASR) represents an extremely challenging task. The major reason is that unwanted additive interference and reverberation are picked up by the microphones besides the desired signal. A hands-free human-machine interface should therefore comprise a powerful acoustic preprocessing unit in line with a robust ASR back-end. However, since perfect speech enhancement cannot be achieved in practice, the output of the front-end will always contain some residual interference and some distortion of the desired signal. It is hence of decisive importance to carefully adjust the hidden Markov models (HMMs) of the ASR system to the front-end. In this contribution, we present a two-channel acoustic front-end based on blind source separation along with Wiener filtering. For the front-end integration into the ASR system, different  types of multi-style as well as adaptive training and HMM adaptation are investigated.