NVIDIA Nemotron 3.5 Transcribe Leads Open Speech AI Benchmarks
NVIDIA Nemotron 3.5 Transcribe unifies multilingual ASR, diarization, translation and audio Q&A, while Granary v2 enables domain adaptation on one GPU.
Summary
NVIDIA released open-weight Nemotron 3.5 Transcribe on October 28, 2026, through Hugging Face and ModelScope. One checkpoint transcribes 40 languages, translates 31 to or from English, attributes up to eight speakers across 13 languages, handles one-hour recordings, answers English audio questions and provides word timestamps. It runs on one NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell systems. An RTX PRO 6000 Blackwell processes 100 audio hours in under three minutes using NVFP4. Deployment options include vLLM, NeMo-Speech.cpp and endpoints from Baseten, DeepInfra and Simplismart, but not live streaming.
Nemotron leads open-weight Hugging Face tests with 5.33% English WER, 4.62% across seven languages and 10.98% long-form WER. It ties Artificial Analysis at 2.8% AA-WER using about 4B active parameters from 30B, versus 24B for its peer, and scores 1.8% on AA-AgentTalk versus 2.3% for the next open model. Its 18.26% diarization error across 21 languages beats Nemotron 3 Diarization’s 18.38%; VoiceArena records 32.4% cpWER, versus 34.2% for Multitalker Parakeet with Nemotron 3 Diarization, and 8.32% deletions against a 31% median. Granary v2 exposes 463,062 hours across 48 languages and 16 corpora, with three transcript forms and reusable NeMo Curator pipelines. On 2,764 held-out Eka Care medical clips, fine-tuning cut WER 45.8%, from 22.44% to 12.17%, after roughly 14 GB300 GPU-hours; keyword WER fell from 9.2% to 6.5% and semantic WER from 10.0% to 7.6%, while LibriSpeech test-clean worsened from 1.54% to 1.69%.
Positives
- Nemotron 3.5 Transcribe covers 40 transcription languages, 31 translation languages and speaker attribution across 13 languages in one open-weight checkpoint.
- 5.33% English WER places the model first among open-weight systems on the Hugging Face Open ASR Leaderboard.
- About 4B active parameters deliver 2.8% AA-WER, tying a model using 24B active parameters.
- 463,062 hours across 48 languages make Granary v2 a substantial reusable resource for speech model adaptation.
- 45.8% lower held-out medical WER shows fine-tuning gains transferred to previously unseen Eka Care speakers.
- 100 hours of audio processed in under three minutes demonstrates high-throughput inference on one RTX PRO 6000 Blackwell GPU.
Risks & concerns
- Nemotron 3.5 Transcribe supports recorded audio rather than live streaming, limiting real-time transcription deployments.
- Audio question answering is trained only in English, and language availability differs across transcription, translation and diarization tasks.
- 18.26% multilingual diarization error and 32.4% VoiceArena cpWER show substantial difficulty remains with speaker attribution and overlapping speech.
- LibriSpeech test-clean WER rose from 1.54% to 1.69% after medical adaptation, indicating a small general-English accuracy trade-off.
- Brand-name drug and diagnostic exact matches still trailed a medical specialist, making real clinician recordings the proposed next step.
- Names, speaker changes and generated answers still require checking against the original recording.