Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model designed for transcription, speaker attribution, translation and audio question answering. Available through Hugging Face and ModelScope, it can process recordings up to an hour long and identify as many as eight speakers.

The model transcribes 40 languages, translates 31 to or from English, and supports speaker attribution in 13. Users can guide its output with natural-language prompts, including requests to label speakers, translate a recording or answer questions about its contents. It also supports word-level timestamps. NVIDIA notes that language coverage varies by task, and that audio question answering was trained in English.

NVIDIA says Nemotron 3.5 Transcribe ranks first among open-weight models on the English Hugging Face Open ASR Leaderboard, with an average word error rate (WER) of 5.33%. The model card also reports the lowest WER among evaluated open-weight systems in its comparisons for multilingual and long-form transcription. On Artificial Analysis’s speech-to-text evaluation, it ties for first among open-weight models at 2.8% AA-WER, using about 4 billion active parameters per token.

The model also ranked first in two speaker-attribution comparisons cited by NVIDIA. On David AI’s 21-language diarization benchmark, it recorded an 18.26% diarization error rate. In VoiceArena’s comparison of 14 systems, it achieved the lowest concatenated minimum-permutation WER, at 32.4%.

Quantized versions are supported for local deployment on systems including RTX Spark, DGX Spark and Jetson AGX Thor, while larger deployments can use Blackwell data-center GPUs. NVIDIA says the model processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU in a benchmark using a forthcoming NVFP4 quantized version. The company describes the model as intended for recorded audio, not live streaming.

Alongside the model, NVIDIA released Granary v2, a multilingual speech-data collection and reusable curation pipeline, with manifests available on Hugging Face and ModelScope. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora. The collection includes multiple transcript formats, as well as labels and data for tasks such as translation and contextual speech recognition. NVIDIA specifies that these figures describe the dataset release, not the complete training mixture for Nemotron 3.5 Transcribe.

Developers can use Granary v2 and NVIDIA NeMo Speech tools to adapt the model to specialist vocabulary and transcript conventions. In a medical speech case study using recordings from Indian doctors and medical students, the best prompt reduced WER on the held-out speakers from 22.44% to 12.17% after fine-tuning, a relative reduction of 45.8%. The study also found a slight increase in WER on LibriSpeech test-clean, from 1.54% to 1.69%.

NVIDIA recommends checking names, speaker changes and generated answers against the original recording. The company provides the model, data manifests and fine-tuning resources through Hugging Face and ModelScope, with serving options including vLLM and SGLang-Omni and hosted endpoints from Baseten, DeepInfra and Simplismart.

Sources: Hugging Face Blog.