Speech to text translation in low resource Setting : LiSTra
In this work we would like investigate the effectiveness of end-to-end architecture in opposite to pipeline approach by using the a transformer like architecture \cite{liu2019synchronous} and a custom attention mechanism that will weight both the speech recognition decoder with the translation decoder.
The used dataset is constructed following instructions in this repository https://github.com/Kabongosalomon/LiSTra.
Credit
- MaSS: A Large and Clean Multilingual Corpus of Sentence-aligned Spoken Utterances Extracted from the Bible
- Yuchen Liu and al. paper
Contact
You can contact me at kabongosalomon@gmail.com