Paper Details

Name: Priyanka Jangra, Prince Kumar, Adeeb Alam, Kriti kant, Nirbhay Mishra & Vikas Babu
Country: India

Year: 2026
Volume: Volume No: 13, January, Year: 2026 (Special Issue)
Page Number: 758-768
DOI: https://doi.org/10.5281/zenodo.19109829

ABSTRACT—
The increasing demand for multilingual communication across the globe highlights the need for effective translation systems, particularly for low-resource languages with scarce or non-existent text data. Conventional speech translation pipelines rely heavily on intermediate text representations, which are impractical for languages with limited written corpora or complex oral traditions. This paper proposes a Textless NLP framework tailored for low-resource speech translation, directly translating speech from source to target languages without the need for text transcription. Leveraging advancements in self-supervised speech representations, speech-only embeddings, and sequence-to-sequence speech mapping, the system captures semantic content from the target language and produces vocal output in the source language. The proposed system is evaluated on simulated low-resource datasets, demonstrating its efficacy in preserving meaning and achieving intelligible translations even in the absence of textual data. This research contributes to the development of inclusive speech technology, particularly for endangered languages, oral dialects, and linguistically marginalized communities. Results indicate that textless speech translation can achieve competitive performance with reduced reliance on annotated parallel corpora, making it a viable solution for real-world deployment in low-resource contexts.

Keywords: Textless NLP, Low-resource languages, Speech translation, Self-supervised learning, Speech-to-speech translation, Endangered languages, Zero-shot translation, Low-resource NLP