You are here
Distant Speech Recognition (DSR) extends the application of Automatic Speech Recognition technology to the very challenging situation that is determined by speaking far from the microphone. An unavoidable drop of ASR performance is introduced in the latter case, due to the resulting mismatch in quality between close-talking and acquired distant speech signals. In fact, a distance of 1-2 meters between speaker and microphones is sufficient to introduce a significant amount of additive (both stationary and non-stationary) background noise and convolutive reverberation effects in the acquired signal. In particular, reverberation regards multi-path propagation in the given acoustic enclosure, and it is well represented by the impulse response between the source position and the microphone position in space. In a real application context both the impulse response and the background noise characteristics are unknown, and this is the main reason of the complexity of this task.
Related keywords: robust acoustic features, acoustic modeling, acoustic impulse response, model adaptation, contaminated speech based training, robust speech recognition, keyword spotting, confidence measures, rejection strategies.