Nvidia has released Nemotron 3.5 Transcribe on Hugging Face and ModelScope, combining speech transcription, speaker attribution, translation and audio question answering in one open-weight checkpoint. It transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13.
The model accepts recordings up to an hour long and can distinguish as many as eight speakers. Natural-language instructions can request a particular format, supply possible specialist terms or ask questions about the audio. Language coverage varies by task, and audio question answering was trained in English. Nvidia advises checking names, speaker changes and generated answers against the original recording.
The release also includes Granary v2 multilingual data, a reusable curation pipeline and NeMo Speech fine-tuning tools. In a medical case study reported by Nvidia, fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt. The company says the base model leads the Hugging Face Open ASR Leaderboard for English, but leaderboard results do not guarantee equal accuracy across accents, domains or recording conditions. Quantized versions for smaller local deployments are planned rather than available in every format today.