Speech Translation on Fisher-Callhome Spanish recipe

This folder contains recipes for tokenization and speech translation with Fisher-Callhome Spanish, a 160-hour Spanish-English ST dataset.

How to run

0- Install extra dependencies

pip install -r extra_requirements.txt

1- Train a tokenizer. The tokenizer takes in input the training translations and determines the subword units that will be used for the ST task, the auxiliary MT task.

cd Tokenizer
python train.py hparams/train_bpe_1k.yaml

2- Train the speech translator

cd ST/transformer
python train.py hparams/transformer.yaml

Performance summary

Results are reported in terms of sacrebleu.

hyperparams file	dev	dev2	test	ctc_weight	asr_weight	mt_weight	Model	GPUs
transformer.yaml	40.67	41.51	40.30	0	0	0	Not Avail.	2xRTX 2080 Ti 11GB
transformer.yaml	47.50	48.33	47.31	1	0.3	0	Model	2xRTX 2080 Ti 11GB
transformer.yaml	46.10	46.56	46.79	1	0.2	0.2	Not Avail.	2xRTX 2080 Ti 11GB
conformer.yaml	46.37	47.07	46.10	0	0	0	Not Avail.	2xRTX 2080 Ti 11GB
conformer.yaml	48.09	48.19	48.04	1	0.3	0	Model	1xTesla A100 (works with 2xRTX 2080 Ti)

About SpeechBrain

• Website: https://speechbrain.github.io/
• Code: https://github.com/speechbrain/speechbrain/
• HuggingFace: https://huggingface.co/speechbrain/

Citing SpeechBrain

Please, cite SpeechBrain if you use it for your research or business.

@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}