At University at Buffalo, I deployed 25 quantized ASR models on a public production web platform, optimizing them for CPU deployment and real-time transcription. I also developed a RAG pipeline for typological feature inference, achieving 72% retrieval accuracy and 70% feature inference accuracy.
At Boston College, I trained and analyzed speech recognition and machine translation models for 30 low-resource languages. Earlier, at Laboratoire d’Informatique de Grenoble, I curated a parallel speech translation dataset for eight languages and trained a visually grounded speech model for image retrieval.
