NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model for multilingual transcription, speaker attribution, translation, and audio Q&A. It supports 40 languages for transcription, 31 for translation, and 13 for speaker diarization, handling up to one hour of audio and eight speakers. Available on Hugging Face and ModelScope, it leads open-weight models on the Hugging Face Open ASR Leaderboard for English and achieves top rankings on speaker attribution benchmarks. Developers can guide output using natural language prompts, adapting to domain-specific vocabulary and accents. Quantized versions are planned for local deployment on RTX Spark, DGX Spark, and Jetson AGX Thor GPUs.
Granary v2, a multilingual data pipeline with fine-tuning recipes, is released alongside the model to extend prompt-level control to training. In a medical case study, fine-tuning reduced word error rate by 45.8% compared to prompt-only adaptation. The model supports word-level timestamps, enabling applications like classroom replay with synchronized transcripts, as demonstrated by Agora. Presets and prompt guides are available in the model card, with task-specific language support varying by function. Audio question answering is currently trained in English only.
Source: Hacker News · Summarized by HeadlinesBriefing