The Pulse
NVIDIA Releases Nemotron 3.5 Transcribe and Granary v2
NVIDIA released Nemotron 3.5 Transcribe, an open-weight model for multilingual transcription, speaker attribution, translation and audio questions. The release also includes Granary v2, a 463,062-hour speech dataset and curation pipeline in
AI.info Team · Written with AI · fact-checked ·
“NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that turns recorded speech into useful text.”
Maryam Motamedi and the other NVIDIA authors
NVIDIA released Nemotron 3.5 Transcribe on October 28, presenting it as a single model for multilingual speech transcription, speaker attribution, translation and audio question answering. Developers can steer its output with natural-language instructions, while a companion release, Granary v2, provides multilingual training data and a reusable pipeline for preparing speech datasets.
The model is open-weight and available through Hugging Face and ModelScope. NVIDIA says it handles recordings up to an hour, supports transcription in 40 languages, translation to or from English in 31, and speaker attribution in 13, with up to eight speakers per recording. The company describes it as a tool for recorded audio, not live-streaming transcription.
One checkpoint, different speech tasks
A prompt can ask the model to transcribe and label speakers, translate speech into English, or answer a question about a meeting. Developers can also specify transcript conventions and provide candidate spellings for specialized terms. NVIDIA’s example prompt for medical speech includes terms such as HbA1c and metformin, while instructing the model to use them only when the audio supports them.
Language coverage depends on the task: the announcement says audio question answering was trained in English. The model also supports word-level timestamps. NVIDIA advises users to check names, speaker changes and generated answers against the recording, a practical limitation for workflows where an incorrect word or attribution could matter.
Benchmark results and the compute claim
NVIDIA reports a 5.33% average word error rate on the English track of Hugging Face’s Open ASR Leaderboard, where the post says it ranks first among open-weight models; API models are excluded from that comparison. The company also cites a 4.62% mean word error rate across seven languages and 10.98% on a long-form track, describing both as the lowest among open-weight models in its comparison.
On Artificial Analysis’s speech-to-text leaderboard, NVIDIA says the model ties for the best open-weight result at 2.8% AA-WER, with about 4 billion active parameters. Its architecture pairs a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model; roughly 4 billion of its 30 billion parameters are active per token. NVIDIA says one RTX PRO 6000 Blackwell Server Edition GPU processed 100 hours of audio in under three minutes using NVFP4 precision, with the measurement based on 30-second clips at high concurrency.
For speaker attribution, the company reports an 18.26% diarization error rate across 21 languages on David AI’s evaluation, and a 32.4% concatenated minimum-permutation word error rate on VoiceArena’s English full-session benchmark. Those results come from separate evaluations with different metrics and language coverage, not a single head-to-head test.
Granary v2 targets domain-specific errors
Granary v2’s public manifest covers 463,062 hours of speech across 48 languages and 16 source corpora. NVIDIA says the dataset is distinct from the full mixture used to train Transcribe. Its pipeline creates multiple transcript forms, including versions that preserve disfluencies and versions that normalize numbers and other written conventions, as well as captions, translation data and speech-grounded question-and-answer pairs.
A medical speech case study offers a more concrete test of adaptation. On 2,764 held-out clips from Indian doctors and medical students, NVIDIA reports that fine-tuning reduced word error rate from 22.44% for the best-prompt baseline to 12.17%, a 45.8% relative reduction. The study also reports a small decline on general English: LibriSpeech test-clean word error rate rose from 1.54% to 1.69% after adaptation.
That trade-off matters: the medical result does not mean domain tuning improves every kind of speech. Developers can download the model and dataset manifests from Hugging Face or ModelScope, serve the checkpoint with vLLM, or run it locally with NeMo-Speech.cpp; NVIDIA also lists hosted endpoints from Baseten, DeepInfra and Simplismart.