Expertise
I work where speech and language AI meets online safety. Each area below gives the problem, what I do about it, and where you can see the evidence. All examples come from my research, not from client work.
Multilingual and low-resource speech and NLP
The problem: most moderation and language tools are built for English and a handful of major languages, while many users speak low-resource languages or mix several in one sentence.
What I do: adapt pre-trained speech and text models (Whisper, wav2vec 2.0 / XLS-R, CLAP, multilingual transformers) to new languages with little labelled data, using few-shot learning, cross-lingual transfer and leave-one-language-out evaluation.
Evidence: few-shot audio abuse detection across 10 Indian languages (COLING 2025); few-shot contrastive adaptation with CLAP (preprint); work under review on how language families shape cross-lingual transfer in speech encoders.
Content moderation and Trust & Safety
The problem: abuse increasingly arrives as voice notes, audio rooms and short videos, and transcribing it first loses tone and fails on informal or profane speech.
What I do: build and evaluate detection pipelines for abusive speech and text, analyse where and why they fail, and prototype monitoring tools quickly.
Evidence: abusive speech detection on the ADIMA benchmark across 10 Indian languages; a racist-narrative tracking pipeline built at the Hack the Feed hackathon.
Responsible AI and bias evaluation
The problem: a model can look accurate on average while failing badly for particular languages or groups, and content itself can carry stereotypes.
What I do: per-language and per-group error audits, bias detection, and LLM-based rewriting checked against human experts.
Evidence: identifying and rectifying gender stereotypes in rhymes and poems (LREC-COLING 2024); ongoing work on how moderation errors are distributed across languages.
Choosing models and architectures
The problem: teams have to decide between transcribe-then-classify pipelines and models that work on the audio directly, and between many possible encoders, often without a fair comparison.
What I do: controlled, budget-matched comparisons with ablations that rule out the obvious confounds, turned into a recommendation that accounts for cost and language coverage.
Evidence: work under review comparing speech-to-text pipelines with direct audio classifiers across 10 Indian languages; in our tests, direct audio models came out ahead.
Generative AI and LLM prototyping
The problem: organisations want to know what LLMs can safely do for them before committing to a build.
What I do: rapid prototypes with LLMs and agents, evaluated against clear baselines and human judgement.
Evidence: LLM rewriting of stereotyped text benchmarked against educators (LREC-COLING 2024); LLM agent tooling for automating my own experiments; hackathon builds.
Digital humanities and media analytics
The problem: media and cultural organisations want to understand language use in content and audiences at scale.
What I do: combine NLP with statistical analysis to study language choice, style and their link to audience success.
Evidence: code-switching in K-pop hits on Billboard charts (CHR 2025); talks including one at TotalEnergies, Paris.
Teaching and training
I teach unsupervised learning sessions in a data science course at Télécom SudParis (teaching), and I enjoy explaining technical results to non-specialist audiences.
Get in touch
If you are working on multilingual moderation, Trust & Safety or Responsible AI, I would be glad to talk. Email me at aditya-narayan.sankaran@telecom-sudparis.eu or connect on LinkedIn.
