Multilingual speech and language AI for safer online spaces

I am a PhD researcher in multilingual speech and language AI at Télécom SudParis, Institut Polytechnique de Paris. I help platforms and organisations detect and audit harmful content across languages, including low-resource and code-mixed ones, using pre-trained speech and text models.

What I can help with

  • Moderation pipelines for many languages: building speech and text classifiers for abusive and hateful content in multilingual and low-resource settings.
  • Choosing the right architecture: deciding whether to transcribe voice content before classifying it, or to work on the audio directly.
  • Fairness and error audits: checking error rates per language and per group, so that a good average score does not hide who bears the errors.
  • New languages with little data: getting a new language working with tens, not thousands, of labelled examples.

See my areas of expertise →

Selected results

  • Few-shot detection of abusive speech across 10 Indian languages, with 50–200 labelled examples per language (COLING 2025).
  • A simple classifier on frozen CLAP audio embeddings comes within 1–3 points of full supervision and far outperforms zero-shot prompting (preprint).
  • A gender-stereotype classifier for rhymes and poems with 97% accuracy, plus LLM rewrites benchmarked against educators’ corrections (LREC-COLING 2024).

Research

My research looks at how pre-trained speech and text representations can help in studying abusive language: what they capture about language, speakers and bias, and how well they transfer to languages they were never trained on. I focus on hate and offensive content moderation in low-resource and code-mixed/code-switched languages, and I also work on computational digital humanities, such as code-switching in K-pop and its link to chart success.

I am supervised by Noel Crespi and Reza Farahbakhsh at the Data Intelligence and Communication Engineering Lab, and I hold a fellowship from the Indo-French Centre for the Promotion of Advanced Research (CEFIPRA/IFCPAR) for the ADEPT project.

Languages: Tamil (native), English (fluent), French (learning).

Get in touch

I am happy to talk about multilingual moderation, Trust & Safety and Responsible AI. Email me at aditya-narayan.sankaran@telecom-sudparis.eu or connect on LinkedIn.

News

[2026-04-20] Our paper on region-grounded 3D PET/CT report generation for Vietnamese, with AI4LIFE at Hanoi University of Science and Technology, will appear at ACL 2026. Read it here.

[2026-04-11] New preprint on few-shot contrastive adaptation (CLAP) for audio abuse detection in low-resource Indic languages. Read it here.

[2026-03-04] Gave a talk at Total Energies, Paris on K-Pop Code-Switching. Find the slides here.

[2025-12-08] Headed to Luxembourg for CHR 2025 to present my work on K-Pop Code-Switching. Find the slides here.

[2025-06-29] I was part of the Summer Institute in Computational Social Science 2025 at ENSAE, Paris.

Older news

[2025-02-01] I officially started my PhD at Télécom SudParis, Institut Polytechnique de Paris.

[2024-11-29] Our work on Few Shot Audio Abuse Detection was accepted as part of COLING 2025.

[2024-05-22] We presented our work on Gender Stereotypes in Rhymes and Poems at LREC-COLING 2024.

[2024-05-02] I received a fellowship under the Indo-French Centre for the Promotion of Advanced Research (CEFIPRA/IFCPAR) for ADEPT.