At Information and Language Processing Research Lab (ILPRL), I developed a Tamang-Nepali-English Tri-Lingual Machine Translation system funded by Google Academic Research under the Society Centered AI category. I also fine-tuned open-source machine translation models and evaluated them on standard metrics.
At ILPRL, I worked on corpus collection and fine-tuning for Newari Automatic Speech Recognition. I also submitted a journal manuscript involving acoustic and linguistic error analysis of Nepali ASR models.
As a Research Intern, I fine-tuned Wav2Vec2, Whisper, and MMS models for Nepali ASR, achieving highly competitive Word Error Rates. I manually collected and curated five hours of Nepali speech corpus for evaluation.
On Shruti, I fine-tuned Wav2Vec2-BERT for Nepali speech recognition and built an inference pipeline with voice activity detection. I designed conversational routing across RAG, general responses, and human escalation, then exposed the workflows through FastAPI with a Streamlit frontend.
