NVIDIA says Nemotron 3.5 Transcribe leads transcription and diarization tests
🎧 Listen to this story
NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that combines transcription, speaker attribution, translation, and audio question answering. The company says it transcribes 40 languages, translates 31 of them to or from English, identifies speakers in 13 languages, handles recordings of up to an hour, and labels up to eight people. The model is available on Hugging Face and ModelScope, with quantized versions planned for local use on devices such as RTX Spark, DGX Spark, and Jetson AGX Thor.
NVIDIA says the model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, has the lowest error rate on David AI’s diarization leaderboard, and placed first in a VoiceArena comparison. The company also introduced the multilingual Granary v2 dataset and pipeline, along with NeMo Speech fine-tuning tools. In a medical study cited by NVIDIA, fine-tuning reduced word error rate by 45.8% relative to the best prompt.
Why it matters · editorial interpretation
By combining transcription, speaker identification, translation, and audio question answering in an open-weight model, Nemotron 3.5 could reduce the need to assemble separate tools for voice-analysis workflows. The accompanying data and fine-tuning tools may also facilitate domain-specific adaptation, although the 45.8% improvement in a medical study is a claim cited by NVIDIA.
Sources
Research