Travis Bartley

Travis Bartley is an NVIDIA deep-learning engineer and speech-technology researcher working on multilingual speech recognition and the company’s Canary and Parakeet model families. His work tackles a central challenge in conversational AI: producing accurate, readable transcripts across languages, dialects, noisy environments, and real-time applications.

A doctoral candidate in English at the City University of New York Graduate Center, Bartley studies multilingualism, dialect, speech recognition, and formal language systems. Earlier, he held Data for Public Good and CUNY Humanities Alliance fellowships and worked as an NVIDIA language-technology research intern. He has also written about public higher education and austerity.

• Language identification: His research on multilingual speech representations found that lower layers of self-supervised Conformer models capture useful signals for identifying languages, including under unfamiliar acoustic conditions.

• Hybrid-autoregressive inference: He coauthored Hybrid-Autoregressive Inference Transducers, which support both autoregressive and non-autoregressive decoding, helping developers balance transcription accuracy against inference speed.

• Readable, context-aware transcripts: His work on punctuated long-sequence training treats punctuation, capitalization, sentence boundaries, and extended context as core speech-recognition concerns; related text-normalization research addresses conversion between written and spoken language.

• Production-scale multilingual models: He coauthored the technical report introducing Canary-1B-v2 and Parakeet-TDT-0.6B-v3, covering 25 primarily European languages and incorporating nonspeech audio to reduce erroneous transcription.

At AI Engineer World’s Fair 2025, Bartley described designing speech systems around noisy audio, dialect and code-switching, specialized vocabulary, and concrete latency requirements.