Model guide

Nemotron 3.5 ASR, explained without the research-paper fog.

Nemotron 3.5 ASR is NVIDIA’s compact streaming speech-recognition model for multilingual transcription. It turns live or recorded audio into punctuated text.

Try the transcriber →

What makes it different

The model is cache-aware: when audio arrives in chunks, it reuses earlier context instead of processing the same overlap again. That reduces latency for live transcription.

  • 0.6B parameters
  • 40 language locales
  • Automatic language detection
  • Native punctuation and capitalization

Best uses

It fits meetings, interviews, voice notes, live captions and voice-agent pipelines where speed matters. Use a dedicated diarization model when you need reliable speaker labels.

Speed versus accuracy

Nemotron supports 1.12s, 0.56s, 0.32s and 0.08s chunk settings. Larger chunks usually favor accuracy; smaller chunks reduce live latency.