artursstaf/speech_translation

Scripts used for Filtering data and Noise Generation as described in the article "Data Augmentation for Pipeline-BasedSpeech Translation" (Baltic HLT 2020).

Speech Translation

Instalation

Clone the repository and create conda environment by running conda env create -f environment.yml.

Synthetic Data Filtering

Filter noise from synthetic speech translation dataset - (source, source_asr, source_pos, target). Source and Source_asr pairs that are not similar are dropped.

python speech_translation.py filter --workdir= --source= --source_asr= --target= --source_pos= --result_source= --result_source_asr --result_target --result_source_pos

Rule-Based Synthetic Noise Generation

Learn suffix noise model from filtered synthetic dataset.

python speech_translation.py learn_noise_model --workdir=workdir --source= --source_pos= --source_asr= --result_model=

Apply suffix noise model to any parallel dataset.

python speech_translation.py apply_noise_model --source= --source_pos= --target= --result_source= --result_target= --noise_model=

Model ASR word split/merges.

python speech_translation.py generate_split_merge --source= --target= --result_target= --result_source=