å Jointly Adversarial Enhancement Training for Robust End-to-End Speech Recognition

This is an implementation of our InterSpeech2019 paper "Jointly Adversarial Enhancement Training for Robust End-to-End Speech Recognition" on Python 3, PyTorch. We propose a jointly adversarial enhancement training to boost robustness of end-to-end systems. Specifically, we use a jointly compositional scheme of mask-based enhancement network, attention-based encoder-decoder network and discriminant network during training.

Requirements

Python 3.5, PyTorch 0.4.0.

Data

AISHELL

You can download AISHELL to run the code. You can run sh run.sh for AISHELL, but it's recommended that you run the commands one by one.

Your Own Dataset

You need build train, dev and test directory. Each has clean_feats.scp noisy_feats.scp and text. You can run python3 data/prepare_feats.py data_dir feat_dir noise_repeat_num to generate the noisy data noisy_feats.scp.

Model

The system consists of an enhancement network, an end-to-end ASR network, a fbank feature extraction network and a discriminant network. The enhancement network transforms the noisy STFT features to the enhanced STFT features. The fbank feature extraction network is used to extract the normalized log fbank features. The end-to-end ASR model estimates the posteriori probabilities for output labels. The discriminant network is used to distinguish between the enhanced features and clean ones.

Training

E2E ASR training

You can train the E2E ASR network using the clean speech data and multi-condition training strategy i.e., optimization with both the clean and noisy speech.

python3 asr_train.py --dataroot Your data directory(including train, dev and test dataset) 

Enhancement Training

You can train the enhancement network by the mask loss function.

python3 enhance_base_train.py --dataroot Your data directory

or the mask fbank loss function.

python3 enhance_fbank_train.py --dataroot Your data directory

or the gan loss function.

python3 enhance_gan_train.py --dataroot Your data directory

Joint Training

You can jointly train the enhancement network and end-to-end ASR network by the ASR loss.

python3 joint_base_train.py --dataroot Your data directory

You can also jointly train the enhancement, end-to-end ASR network and a discriminant network by the adversarial loss.

python3 joint_train.py --dataroot Your data directory

Decoding

We use the beam search for decoding in all the experiments.

python3 asr_recog.py