NVIDIA Releases Open Weight Nemotron 3.5 Transcribe 30B MoE Speech Model Ranking First on Voice Arena Diarization
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵
Context
Nemotron 3.5 Transcribe packs 40 language speech transcription, multi speaker attribution, translation, contextual prompting, and audio understanding into an open weight 30B Mixture of Experts checkpoint with only 4B active parameters, taking first place on Voice Arena's diarization leaderboard.
Speech to text and speaker diarization have traditionally required chaining separate acoustic transcription, timestamp alignment, and speaker clustering pipelines. Earlier on Wednesday, Mistral AI highlighted that its proprietary Voxtral Mini Transcribe 3 ranked second on Voice Arena's joint diarization and automatic speech recognition leaderboard.
Deep dive
NVIDIA AI released Nemotron 3.5 Transcribe an open weight, promptable 30 billion parameter Mixture of Experts speech model that activates roughly 4 billion parameters per token. Instead of treating automatic speech recognition, speaker diarization, translation, and audio instruction following as separate stages, the model unifies them inside a single checkpoint covering 40 languages. Users can pass an audio recording alongside natural language instructions asking the model to label speakers during rapid overlapping talk, translate the transcript into Spanish or Japanese, or locate and redact personally identifiable information.
Across external speech benchmarks, Nemotron 3.5 Transcribe ranks number one on Voice Arena's 14 system speaker diarization and automatic speech recognition comparison, placing directly ahead of Mistral AI's Voxtral Mini Transcribe 3, and also takes first place on David AI's multilingual diarization benchmark. On the Artificial Analysis non streaming speech to text leaderboard, Nemotron 3.5 Transcribe achieves a 2.81 percent word error rate with 4 billion active parameters, nearly matching the 2.77 percent word error rate of the 24 billion parameter dense Voxtral Small model at roughly one sixth the active compute.
Alongside the model weights, NVIDIA released the Granary v2 open speech dataset and NeMo Speech fine tuning recipes on Hugging Face. In a medical automatic speech recognition case study shared by the team, domain fine tuning with NeMo Speech cut word error rate by 45.8 percent relative to the best zero shot prompt on held out clinical speakers.
What to expect next
Developers can download the open model weights, training data, and NeMo Speech fine tuning recipes from Hugging Face today or test custom recordings in the hosted Hugging Face Space. Because the 30B Mixture of Experts architecture activates only 4B parameters per forward pass, enterprise voice platforms, contact centers, and clinical scribes can self host state of the art multilingual diarization and PII redaction at low inference cost.