tuanh23 / multimodalst

Limit the use of end-to-end data for Speech Translation (by leveraging Automatic Speech Recognition and Machine Translation data instead) using zero-shot multilingual text translation techniques.

Tackling data scarcity in Speech Translation using zero-shot multilingual Machine Translation techniques

This repository is derived from the NMTGMinor project at https://github.com/quanpn90/NMTGMinor
The SVCCA calculation is derived from https://github.com/nlp-dke/svcca

Powered by Mediaan.com

Speech Translation (ST) is the task of translating speech audio in a source language into text in a target language. This repository implements and experiments on different approaches for ST:

• Cascaded ST, including 2 steps: Automatic Speech Recognition (ASR) and Machine Translation (MT)
• Direct ST: models trained only on ST data
• (Main contribution) End-to-end ST limiting the use of ST data: multi-modal models leveraging ASR and MT training data for ST task

The Transformer architecture is used as the baseline for the implementation.

High-level instruction to use the repo:

• Run covost_data_preparation.py to download and preprocess the data.
• Run the shell script of interst, change the variables in the script if needed.
  • run_translation_pipeline.sh for single-task models (ASR, MT, ST)
  • cascaded_ST_evaluation.sh evaluates cascaded ST using pretrained ASR and MT models
  • run_translation_multi_modalities_pipeline.sh for multi-task, multi-modality models (including zero-shot)
  • run_zeroshot_with_artificial_data.sh for zero-shot models using data augmentation
  • run_bidirectional_zeroshot.sh for zero-shot models using additional opposite training data
  • run_fine_tunning.sh, run_fine_tunning_fromASR.sh for fine-tuning models with ST data, resulting in few-shot models
  • modality_similarity_svcca.sh, modality_similarity_classifier.sh measure text-audio similarity in representation

See notebooks/Repo_Instruction.ipynb for more details.

Citation

@INPROCEEDINGS{9746815,
  author={Dinh, Tu Anh and Liu, Danni and Niehues, Jan},
  booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Tackling Data Scarcity in Speech Translation Using Zero-Shot Multilingual Machine Translation Techniques}, 
  year={2022},
  volume={},
  number={},
  pages={6222-6226},
  doi={10.1109/ICASSP43922.2022.9746815}}