Detail publikačního výsledku

End-to-End Speech Translation for Low-Resource Languages Using Weakly Labeled Data

POTHULA, A.; AKKIRAJU, B.; BANDARUPALLI, S.; D, C.; KESIRAJU, S.; VUPPALA, A.

Originální název

End-to-End Speech Translation for Low-Resource Languages Using Weakly Labeled Data

Anglický název

End-to-End Speech Translation for Low-Resource Languages Using Weakly Labeled Data

Druh

Stať ve sborníku v databázi WoS či Scopus

Originální abstrakt

The scarcity of high-quality annotated data presents a significant challenge in developing effective end-to-end speech-to-text translation (ST) systems, particularly for low-resource languages. This paper explores the hypothesis that weakly labeled data can be used to build ST models for low-resource language pairs. We constructed speech-to-text translation datasets with the help of bitext mining using state-of-the-art sentence encoders. We mined the multilingual Shrutilipi corpus to build Shrutilipi-anuvaad, a dataset comprising ST data for language pairs Bengali-Hindi, Malayalam-Hindi, Odia-Hindi, and Telugu-Hindi. We created multiple versions of training data with varying degrees of quality and quantity to investigate the effect of quality versus quantity of weakly labeled data on ST model performance. Results demonstrate that ST systems can be built using weakly labeled data, with performance comparable to massive multi-modal multilingual baselines such as SONAR and SeamlessM4T.

Anglický abstrakt

The scarcity of high-quality annotated data presents a significant challenge in developing effective end-to-end speech-to-text translation (ST) systems, particularly for low-resource languages. This paper explores the hypothesis that weakly labeled data can be used to build ST models for low-resource language pairs. We constructed speech-to-text translation datasets with the help of bitext mining using state-of-the-art sentence encoders. We mined the multilingual Shrutilipi corpus to build Shrutilipi-anuvaad, a dataset comprising ST data for language pairs Bengali-Hindi, Malayalam-Hindi, Odia-Hindi, and Telugu-Hindi. We created multiple versions of training data with varying degrees of quality and quantity to investigate the effect of quality versus quantity of weakly labeled data on ST model performance. Results demonstrate that ST systems can be built using weakly labeled data, with performance comparable to massive multi-modal multilingual baselines such as SONAR and SeamlessM4T.

Klíčová slova

weakly labeled data, speech translation, end-to- end models, low-resource languages

Klíčová slova v angličtině

weakly labeled data, speech translation, end-to- end models, low-resource languages

Autoři

POTHULA, A.; AKKIRAJU, B.; BANDARUPALLI, S.; D, C.; KESIRAJU, S.; VUPPALA, A.

Rok RIV

2026

Vydáno

17.08.2025

Nakladatel

ISCA

Místo

Rotterdam

Kniha

Interspeech 2025

Periodikum

Interspeech

Stát

Nizozemsko

Strany od

41

Strany do

45

Strany počet

5

URL

https://www.isca-archive.org/interspeech_2025/pothula25_interspeech.pdf

BibTex

@inproceedings{BUT199932,
  author="{} and  {} and  {} and  {} and Santosh {Kesiraju} and  {}",
  title="End-to-End Speech Translation for Low-Resource Languages Using Weakly Labeled Data",
  booktitle="Interspeech 2025",
  year="2025",
  journal="Interspeech",
  pages="41--45",
  publisher="ISCA",
  address="Rotterdam",
  doi="10.21437/interspeech.2025-2525",
  url="https://www.isca-archive.org/interspeech_2025/pothula25_interspeech.pdf"
}