NVIDIA releases open-weight Nemotron 3.5 Transcribe, which ties for #1 among open-weight speech-to-text models on Artificial Analysis

NVIDIA's Nemotron 3.5 Transcribe is a new open-weight, promptable speech-to-text model: a 30B mixture-of-experts with about 4B active parameters. NVIDIA says one checkpoint handles transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding, and that the weights, data and recipes are open. It is available on Hugging Face, with a live demo.

Results Artificial Analysis reports (its own figures):

• It scores 2.8% on the AA-WER Index, tied with Mistral's Voxtral Small for #1 among open-weight models and ahead of Thinking Machines' Inkling (3.5%). It improves on NVIDIA's previous best, Canary Qwen 2.5B (4.3%).
• On AA-AgentTalk, its dataset of voice agent conversations, it scores 1.8% against 2.3% for Voxtral Small.
• It matches Voxtral Small's accuracy with 6x fewer active parameters (4B against 24B); other open-weight models of a similar size sit around 4% WER.
• On DeepInfra's hosted endpoint it runs at a Speed Factor of 75x (a minute of audio in under a second) and costs $2.00 per 1,000 minutes of audio.

Claims NVIDIA makes:

• It ranks #1 among open-weight models on the Hugging Face Open ASR English leaderboard, with a 5.33% average word error rate.
• It ranks #1 in VoiceArena's 14-system diarization-plus-ASR comparison and on David AI's multilingual diarization benchmark.
• In a medical ASR case study, fine-tuning with Granary v2 open speech data and NeMo Speech tools cut word error rate by 45.8% relative to the best prompt on held-out speakers.

The Artificial Analysis post had 64 points on X by October 28; NVIDIA's announcement had 6,612 points and 230 comments by October 29.