NVIDIA Nemotron 3.5 Transcribe: One Promptable Model for Multilingual Speech Intelligence

NVIDIA has unveiled Nemotron 3.5 Transcribe, a single open-weight checkpoint that handles multilingual transcription, speaker attribution, translation, and audio question answering — all steerable through natural-language prompts. The model transcribes 40 languages, translates 31 to or from English, and performs diarization for up to eight speakers across 13 languages, processing recordings up to one hour long.

Benchmark results place it first among open-weight models on the Hugging Face Open ASR Leaderboard for English, atop Artificial Analysis's accuracy–active-parameter frontier, and with the lowest diarization error rate on David AI's 21-language leaderboard. VoiceArena's 14-system comparison also ranks it first. Critically, it runs on a single NVIDIA GPU, scaling from DGX Spark workstations to Blackwell data-center systems.

Granary v2: From Prompting to Fine-Tuning

Complementing the model, Granary v2 releases the multilingual training corpus, a reproducible curation pipeline, and NeMo Speech fine-tuning recipes. This lets teams move beyond prompt engineering into weight-level adaptation. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt-only baseline — demonstrating how domain-specific vocabulary and accent adaptation translate to measurable accuracy gains.

For AI infrastructure teams, the takeaway is clear: a single versatile checkpoint reduces model sprawl, while GPU-efficient inference keeps compute costs predictable. The open weights and data pipeline also eliminate vendor lock-in, letting organizations deploy on-premises or in any cloud — including AiGpu's GPU cloud — with full control over data residency and latency.