NVIDIA Nemotron 3.5 Transcribe 30B A4B
What it is
An open-weight, offline speech and audio model designed for multilingual transcription, speech translation, and speaker diarization. It utilizes a hybrid mixture-of-experts architecture with 34B total parameters, activating approximately 4.3B parameters per inference to support recordings up to one hour long across 40 languages.
The problem it solves
The high computational cost and limited language support of traditional large-scale ASR models, as well as the lack of open-weight models capable of handling long-form audio with speaker attribution and contextual understanding.
What you could use it for
- Transcribing long-form audio recordings in 40 languages
- Speaker-attributed transcription and diarization
- Speech translation and contextual transcription
- Audio question answering and instruction-following
Why it matters
Enables high-accuracy, low-latency multilingual speech processing for enterprise and research applications by providing an efficient, open-weight model that balances quality with computational cost via its MoE architecture.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.