NVIDIA Nemotron 3.5 ASR Streaming 0.6B
NVIDIA's 0.6B streaming ASR model covering 35 languages.
Last verified:
Visit NVIDIA Nemotron 3.5 ASR Streaming 0.6B
What is NVIDIA Nemotron 3.5 ASR Streaming 0.6B?
NVIDIA Nemotron 3.5 ASR Streaming 0.6B is a multilingual, streaming automatic speech recognition model built on a Cache-Aware FastConformer-RNNT architecture with 600M parameters. It transcribes speech across 40 language-locales from a single model via language-ID prompt conditioning, targeting low-latency voice agents and high-throughput batch transcription workloads.
NVIDIA Nemotron 3.5 ASR Streaming 0.6B pricing
Pricing model: Free
Free open-weights model released under the OpenMDW-1.1 license (inference infrastructure costs apply)
NVIDIA Nemotron 3.5 ASR Streaming 0.6B pros
- Single model covers 40 language-locales with language-ID prompting and optional automatic language detection
- Cache-aware architecture avoids redundant computation, improving throughput and reducing latency versus buffered streaming
- Configurable chunk sizes (80ms–1120ms) let users trade latency vs. accuracy at inference without retraining
- Built-in punctuation and capitalization; ready for commercial use under OpenMDW-1.1 license
NVIDIA Nemotron 3.5 ASR Streaming 0.6B cons
- Only 19 of 40 locales are transcription-ready; 13 are broad-coverage and 8 require fine-tuning on in-domain data
- 600M parameter size may be heavier than purpose-built single-language models
- Optimized for NVIDIA GPUs (e.g., H100), so deployment is tied to NVIDIA hardware
Frequently asked questions about NVIDIA Nemotron 3.5 ASR Streaming 0.6B
How many languages does Nemotron 3.5 ASR support?
It supports 40 language-locales: 19 transcription-ready, 13 broad-coverage, and 8 adaptation-ready that require fine-tuning.
Can the model handle both streaming and batch transcription?
Yes, it is designed for both low-latency streaming and high-throughput batch workloads, with configurable chunk sizes from 80ms to 1120ms.
Is the model available for commercial use?
Yes, NVIDIA states the model is ready for commercial use and is released under the OpenMDW-1.1 license.
Can the model be fine-tuned for better accuracy on specific languages?
Yes, NVIDIA provides a blog post describing how to fine-tune Nemotron 3.5 ASR with before/after results.