Details
- NVIDIA introduced Nemotron 3.5 Transcribe, a 30-billion-parameter mixture-of-experts model with approximately 4.3 billion active parameters.
- The model combines transcription, speaker attribution, translation, contextual prompting and audio question answering in one checkpoint.
- It supports 40 languages, recordings up to one hour and a 65,536-token context window, with optional word timestamps.
- NVIDIA says the model ranks first among open-weight systems on the Hugging Face Open ASR English leaderboard with a 5.33% average word error rate.
- The model also ranked first in VoiceArena’s 14-system diarization and ASR comparison and in David AI’s multilingual diarization benchmark, according to NVIDIA.
- NVIDIA reported a 2.81% word error rate on an Artificial Analysis benchmark, close to Voxtral Small’s 2.77% result despite using roughly one-sixth as many active parameters.
- Developers can fine-tune the open weights with Granary v2 speech data and NeMo Speech tools; a medical ASR case study reduced word error rate by 45.8% relative to the best prompt.
- Weights, data and training recipes are available on Hugging Face, with deployment options including NVIDIA RTX Spark and DGX Spark systems, Jetson AGX Thor and GeForce RTX 5090 GPUs.
Impact
Nemotron 3.5 Transcribe strengthens NVIDIA’s push to make capable speech models deployable locally rather than through per-call APIs. Its combination of multilingual transcription, diarization and translation reduces the need to assemble several systems, while the relatively low active-parameter count could lower hardware requirements. The open weights and fine-tuning recipes also give enterprises more control over sensitive recordings, although real-world accuracy, licensing and deployment costs will determine how much pressure it places on established speech platforms.