ML Engineer | Speech & NLP for low-resource languages
I build ML systems that ship, and I do research on speech and language technology for languages that most models ignore. Master's in Data Science at Duke; previously, four years at Ishango.ai and a research intern at Mila.
📌 Looking for a summer 2027 internship in ML, speech, or applied AI. Reach me at luelbhagos@gmail.com.
Portfolio • LinkedIn • Hugging Face • Google Scholar
- Tigrinya GPT-2: 51M-parameter model trained from scratch on 21M Tigrinya tokens. Demo · Model
- Gemma-3-4B Tigrinya: SFT and instruction-tuned Gemma 3 for Tigrinya. Base · QA
- mSTEB: multilingual benchmark of LLMs on speech and text tasks. Paper · Leaderboard
- Ge'ez TTS: end-to-end text-to-speech for Ge'ez with Tacotron 2 + WaveGlow, including text normalization and alignment.
- mSTEB: Massively Multilingual Evaluation of LLMs on Speech and Text Tasks · IEEE ASRU 2025 (co-first author)
- EmoFI-TTS: Face-Conditioned Expressive Voice Synthesis with Intensity Modulation · under review (co-first author)
- Ibom NLP: A Step Toward Inclusive NLP for Nigeria's Minority Languages · IJCNLP-AACL 2025
- OpenBibleTTS: Large-Scale Speech Resources and TTS Models for Low-Resource Languages · under review
