PhD Researcher · Speech-to-Speech Translation · Dakar, Senegal
I build speech and language technology for low-resource African languages — with a focus on automatic speech recognition, neural machine translation, and self-supervised learning for speech.
Currently a PhD researcher at Université Cheikh Anta Diop (UCAD), working on end-to-end Speech-to-Speech Translation for Wolof, Pulaar, and other West African languages. Previously Lead Data Scientist at LAfricaMobile, where I led a team building NLU/NLG and end-to-end TTS systems for local languages.
Open to research collaborations and roles building production speech systems for under-served languages.
Université Cheikh Anta Diop · Dakar, Senegal
LAfricaMobile · Dakar, Senegal
Omdena · Remote
Various roles · Senegal
The first cross-lingual open-retrieval QA dataset for 10 African languages — establishing benchmarks for multilingual retrieval and reader systems on truly low-resource languages.
Comparing direct and cascaded S2ST architectures for Wolof–French and Pulaar–French, with a focus on data-efficient pre-training strategies and evaluation under limited parallel speech.
Empirical study of Whisper fine-tuning regimes for Wolof — examining data scale, augmentation, and dialectal generalization. Released open checkpoints and a 120h transcribed Wolof corpus.
Proof-of-concept on knowledge distillation from a frozen BERT teacher into a Wav2Vec2 student — exploring whether textual contextual representations can guide acoustic self-supervised learning when paired data is scarce.
Master's thesis on rank-metric codes and their application to post-quantum encryption and digital signature schemes.
An NLP toolkit for the Wolof language: text classification, neural machine translation, and automatic speech recognition. Designed to be simple, easy to use, and a starting point for low-resource African NLP.
A voice assistant in Wolof, in the spirit of Siri / Alexa, with built-in code-switching detection and correction. Designed to make digital tools accessible — including for non-literate users — in agriculture, fishing, education, and healthcare.
An implementation study of direct speech-to-speech translation with a sequence-to-sequence model — re-creating the Translatotron architecture as a foundation for African-language S2ST experiments.
A study of Neural Audio Codecs as discrete representations for ASR — investigating whether quantized codec tokens (à la EnCodec / SoundStream) can serve as efficient inputs for downstream speech recognition.
Whisper · Wav2Vec2 · HuBERT · ASR · TTS · S2ST · Self-supervised pre-training · Neural audio codecs
Transformers · NMT · Tokenization · Language modelling · Multilingual evaluation · Low-resource methods
Python · PyTorch · 🤗 Transformers / Datasets · FastAPI · Docker · MLflow · W&B · AWS
Mathematics · Coding theory · Statistical learning · Information systems · Software engineering
Wolof · French · English · Arabic (intermediate)
Machine Learning · Deep Learning · NLP · TA at UCAD