AI BriefingNVIDIA Nemotron 3.5 Transcribe Released as Open Source
·2026.11.02 09:00
Key point
NVIDIA has released Nemotron 3.5 Transcribe, which supports 40 languages, as open source.
Details
- Performs ASR for 40 languages, translation for 31 languages, and speaker diarization for up to 8 speakers with a single model
- Achieved the top rank among open-weight models on the Hugging Face Open ASR leaderboard with an average English WER of 5.33%
- Simultaneously released the Granary v2 dataset (48 languages, 463,062 hours) and the NeMo Curator pipeline
- Consumer GPUs such as the RTX 5090 support only NVFP4 quantization, with an average error rate 3.6% higher than BF16
Execution Environment and Serving
- The combination of vLLM 0.32.0 and the NeMo Speech plugin supports all features, including timestamps and speaker diarization
- SGLang-Omni 0.2.0 supports only file-based ASR and does not support speaker diarization or translation features
- Official examples are provided based on a single 80GB H100; environment variable configuration is required when processing 1 hour of recordings
Performance and Benchmarks
- Recorded a CER of 2.78% for Korean ASR on FLEURS and 100% language identification accuracy
- Achieved an audio processing speed of up to 2,238 seconds of audio per second in a single GB300 NVFP4 environment
- Provides a DeepInfra hosted API with a pricing plan of $0.002 per minute of audio ($0.12 per hour)
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.