InfoReleasesHacker News · community
Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages
NVIDIA released Nemotron 3.5 Transcribe, an open-weight promptable speech model handling transcription in 40 languages, translation for 31, and speaker attribution in 13, with up to one-hour recordings and eight speakers. It tops open-weight benchmarks (5.33% English WER, 2.8% AA-WER with ~4B active parameters) and is available on Hugging Face and ModelScope, servable via vLLM or hosted endpoints from Baseten, DeepInfra, or Simplismart. The open Granary v2 data pipeline (463,062 hours across 48 languages) and NeMo Speech fine-tuning tools let developers adapt the model to domain vocabulary, w…
Read the full story on huggingface.co
What to doTry the checkpoint via vLLM, SGLang-Omni, or a hosted endpoint, and use Granary v2 manifests plus NeMo Speech recipes if you need domain-specific adaptation.
- Source
- huggingface.co