statnlp / ada4asr

Implementation of our paper: On-the-Fly Aligned Data Augmentation for Sequence-to-Sequence ASR

Requirements

• Python 3.6
• fairseqv1.0

Scripts added/modified for ADA

• fairseq/criterions/s2t_xent_ctc_loss.py
• fairseq/criterions/label_smoothed_cross_entropy.py
• fairseq/models/speech_to_text/init.py
• fairseq/models/speech_to_text/s2t_transformer.py
• fairseq/models/speech_to_text/s2t_ctc_transformer.py
• fairseq/models/roberta/hub_interface.py
• fairseq/models/transformer.py
• fairseq/data/audio/speech_to_text_dataset.py
• fairseq/data/audio/audio_dict_dataset.py
• fairseq/data/audio/feature_transforms/alignAugment.py
• fairseq/data/audio/feature_transforms/samp_fbank.py
• fairseq/data/audio/feature_transforms/specaugment.py
• fairseq/tasks/speech_to_text.py
• fairseq/sequence_generator.py

sample data input (.tsv)

script for audio dictionary