Neural network based GEV beamformer

Introduction

This repository contains code to replicate the results for the 3rd CHiME challenge using a NN-GEV Beamformer.

Install

This code requires Python 3 to run (although most parts should be compatible with Python 2.7). Install the necessary modules:

pip install chainer
pip install tqdm
pip install SciPy
pip install scikit-learn
pip install librosa

Usage

1. Extract the speech and noise images for the SimData using the modified Matlab script in CHiME3/tools/simulation

2. Start the training for the BLSTM model using the GPU with id 0 and the data directory data:

   python train.py --chime_dir=../chime/data --gpu 0 data BLSTM
   

   This will first create the training data (i.e. the binary mask targets) and then run the training with early stopping. Instead of BLSTM it is also possible to specify FW to train a simple feed-forward model.

3. Start the beamforming:

```
beamform.sh ../chime/data data/export_BLSTM data/BLSTM_model/best.nnet BLSTM
```

This will apply the beamformer to every utterance of the CHiME database and store the resulting audio file in ``data/export_BLSTM``. The model ``data/BLSTM_model/best.nnet`` is used to generate the masks.

4. Start the kaldi baseline using the exported data.

If you want to use the beamformer with a different database, take a look at beamform.py and chime_data and modify it accordingly.

Results

With the new baseline, you should get the following results:

```
local/chime4_calc_wers.sh exp/tri3b_tr05_multi_noisy new_baseline exp/tri3b_tr05_multi_noisy/graph_tgpr_5k
compute dt05 WER for each location
-------------------
best overall dt05 WER 9.77% (language model weight = 12)
-------------------
dt05_simu WER: 9.81% (Average), 8.95% (BUS), 11.28% (CAFE), 8.55% (PEDESTRIAN), 10.44% (STREET)
-------------------
dt05_real WER: 9.73% (Average), 11.67% (BUS), 9.37% (CAFE), 8.41% (PEDESTRIAN), 9.47% (STREET)
-------------------
et05_simu WER: 10.67% (Average), 8.85% (BUS), 11.34% (CAFE), 11.02% (PEDESTRIAN), 11.47% (STREET)
-------------------
et05_real WER: 14.00% (Average), 19.01% (BUS), 13.37% (CAFE), 12.37% (PEDESTRIAN), 11.24% (STREET)
-------------------


./local/chime4_calc_wers_smbr.sh exp/tri4a_dnn_tr05_multi_noisy_smbr_i1lats new_baseline exp/tri4a_dnn_tr05_multi_noisy/graph_tgpr_5k
compute dt05 WER for each location
-------------------
best overall dt05 WER 5.87% (language model weight = 9) (Number of iterations = 4)
-------------------
dt05_simu WER: 5.62% (Average), 5.24% (BUS), 6.58% (CAFE), 4.91% (PEDESTRIAN), 5.77% (STREET)
-------------------
dt05_real WER: 6.11% (Average), 7.66% (BUS), 5.83% (CAFE), 5.10% (PEDESTRIAN), 5.87% (STREET)
-------------------
et05_simu WER: 7.26% (Average), 6.74% (BUS), 7.70% (CAFE), 7.38% (PEDESTRIAN), 7.23% (STREET)
-------------------
et05_real WER: 9.48% (Average), 14.06% (BUS), 8.22% (CAFE), 7.81% (PEDESTRIAN), 7.84% (STREET)
-------------------


local/chime4_calc_wers.sh exp/tri4a_dnn_tr05_multi_noisy_smbr_lmrescore new_baseline_rnnlm_5k_h300_w0.5_n100 exp/tri4a_dnn_tr05_multi_noisy_smbr_lmrescore/graph_tgpr_5k
compute dt05 WER for each location
-------------------
best overall dt05 WER 4.02% (language model weight = 11)
-------------------
dt05_simu WER: 3.97% (Average), 3.66% (BUS), 4.65% (CAFE), 3.38% (PEDESTRIAN), 4.19% (STREET)
-------------------
dt05_real WER: 4.07% (Average), 5.34% (BUS), 3.61% (CAFE), 3.35% (PEDESTRIAN), 4.00% (STREET)
-------------------
et05_simu WER: 4.51% (Average), 4.09% (BUS), 4.61% (CAFE), 4.46% (PEDESTRIAN), 4.86% (STREET)
-------------------
et05_real WER: 6.46% (Average), 9.87% (BUS), 5.47% (CAFE), 5.14% (PEDESTRIAN), 5.34% (STREET)
-------------------
```

Citation

If you use this code for your experiments, please consider citing the following paper:

@inproceedings{Hey2016,
title = {NEURAL NETWORK BASED SPECTRAL MASK ESTIMATION FOR ACOUSTIC BEAMFORMING},
author = {J. Heymann, L. Drude, R. Haeb-Umbach},
year = {2016},
date = {2016-03-20},
booktitle = {Proc. IEEE Intl. Conf. on Acoustics, Speech and Signal Processing (ICASSP)},
keywords = {},
pubstate = {forthcoming},
tppubtype = {inproceedings}
}