Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages

NVIDIA Nemotron 3.5 Transcribe is an open-weight model for speech transcription, speaker attribution, translation, and audio question answering. Natural-language prompts can shape transcripts for different uses, such as labeling speakers, translating speech, answering questions, or preserving domain-specific terminology. It supports 40 transcription languages, translation to or from English for 31 languages, and speaker attribution in 13 languages; recordings can run up to an hour and include as many as eight speakers. The model supports word-level timestamps and overlapping speech, but is designed for recorded audio rather than live streaming. It ranks first among open-weight models on the Hugging Face English ASR leaderboard, with 5.33% average word error rate, and reaches 2.8% on Artificial Analysis’s accuracy measure with about 4 billion active parameters. It also leads evaluated systems on two speaker-attribution benchmarks.

For teams adapting speech recognition to specialized vocabulary, accents, and transcript conventions, the release includes Granary v2, an open data-curation pipeline and multilingual dataset, plus NeMo Speech fine-tuning tools. Granary v2’s public manifest covers 463,062 hours across 48 languages and 16 source corpora; these figures describe the released data, not the model’s entire training mixture. In a medical case study, fine-tuning reduced word error rate on held-out speakers by 45.8% compared with the best prompt-based result. The model runs on one NVIDIA GPU, from edge hardware to data-center systems, and can be served with vLLM or run locally with NeMo-Speech.cpp.