asr1

End-to-End Far-Field Speech Recognition with Unified Dereverberation and Beamforming

This recipe provides source code for data simulation and models related to the paper "End-to-End Far-Field Speech Recognition with Unified Dereverberation and Beamforming".

Prerequisite Installation

1. Kaldi and ESPnet (v.0.5.3 in this repository)
2. MATLAB (for data simulation)

Steps

1. This recipe uses spatialized wsj1-2mix as the dataset, and the data simulation scripts and instructions can be found in data_simu/.

2. After generating the spatialized wsj1-2mix data (16 kHz, max version), just run run.sh to start data preparation, training and evaluation. You can specify some arguments to control which stages to run, e.g.

./run.sh --stage 4 --stop-stage 5

Note: You may need to modify the default paths specified in run.sh to make it work.

Note

1. The implementation of frontend modules (both WPE+MVDR and WPD) can be found in https://github.com/Emrys365/espnet/blob/wsj1_mix_spatialized/espnet/nets/pytorch_backend/frontends/frontend.py and https://github.com/Emrys365/espnet/blob/wsj1_mix_spatialized/espnet/nets/pytorch_backend/frontends/frontend_wpd.py respectively.