What changed
NVIDIA has announced the release of Nemotron 3.5 Transcribe and Granary v2, an open-weight model for multilingual transcription, speaker attribution, translation, and audio question answering. The model is now available on Hugging Face and ModelScope. It supports transcription in 40 languages, translation in 31 languages, and speaker attribution in 13 languages. The model can handle audio recordings up to an hour long and label up to eight speakers. Developers can guide the output with natural language instructions and adapt the model to domain-specific vocabulary and accents.
What this means in practice
The model is designed to provide accurate transcription, speaker attribution, and translation for various use cases, including multilingual calls, fast-paced meetings, and medical dictation. It is the first open-weight model to rank first on the Hugging Face Open ASR Leaderboard for English and leads in several benchmark tests for speaker attribution. The model also includes a data pipeline and fine-tuning tools for developers to adapt the model to specific applications. The Granary v2 data pipeline and NeMo Speech fine-tuning tools allow developers to define the written form they need and adapt the model to the speech their applications encounter.
Limitations and availability
Quantized versions of the model are supported, allowing it to run locally on devices like RTX Spark, DGX Spark, Jetson AGX Thor, or scale up on Blackwell data-center GPUs. The model card includes recommended presets, a full prompt guide, and a task-by-language matrix, noting that language support differs by task and audio question answering is trained in English. Word-level timestamps are also supported. The model is available on Hugging Face and ModelScope, with the data pipeline and fine-tuning tools also accessible on these platforms.
