NVIDIA has released Nemotron 3.5 Transcribe, a free artificial intelligence model designed to convert speech into text across 40 languages. The system identifies up to 8 distinct speakers per audio file and answers direct questions about the recorded conversation. While it tops the Open ASR Leaderboard on Hugging Face, the model does not support real time processing.
Nemotron 3.5 Transcribe processes recordings up to 1 hour in length. The model generates verbatim or edited transcripts with precise timestamps, translates spoken audio, and extracts specific answers based on audio content. All operations occur in batch mode, making the software unsuitable for live streaming or immediate dictation tasks.
NVIDIA built the model with 30 billion parameters, optimizing the code to run efficiently on a single chip. It scales from the Jetson AGX Thor robotics module and the DGX Spark workstation up to enterprise Blackwell silicon in data centers. On the Hugging Face Open ASR Leaderboard, the software outperforms competing speech recognition tools in transcription accuracy tests.
The open weights are available as a free download on the Hugging Face platform. Developers can adapt the model for custom applications through the NeMo Speech software toolkit. Alongside the model release, NVIDIA published Granary v2, an expanded speech training library containing 463,062 hours of audio across 48 languages.
