Room Impulse Response Estimation from Speech Signal

When travelling in a room, any sound is distorted by a room impulse response (RIR). Determining RIR has always been an important task in acoustics, but nowadays, it is even more important, as RIR can be used to augment data for training automatic speech recognition (ASR) systems. Classically, a RIR is estimated from a pair of clean and reverberated sound signals. This is however not practical for real scenarios (such as personal assistants, smart homes, etc.), as the clean signal is not available. The aim of the bachelor thesis is to investigate ''blind'' RIR estimation only from a reverberated speech signal. We have used the BUT ReverbDB data set and first, re-implemented techniques for classical clean-reverberated signals estimation of RIRs. Then, we investigated two techniques for RIR estimation only from a reverberated signal. The first technique uses reverberated impulse-like phonemes in speech which are expected to resemble RIR. Averaging and deconvolution of these phonemes were tested to improve the quality and robustness of the estimation. The second technique makes use of a regression neural networks trained to produce the RIR from a speech input. Although none of the techniques reaches the quality of classical measurement, the estimated RIRs have the potential to help in augmenting data for ASR system training.

Room Impulse Response Estimation, Speech signal