Trending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#8Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#9krea2-identity-editconradlocke⬇0kTrending Model:#10Laguna-S-2.1poolside⬇3kTrending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#8Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#9krea2-identity-editconradlocke⬇0kTrending Model:#10Laguna-S-2.1poolside⬇3k

NVIDIA Drops Nemotron-3.5-Asr-Streaming-0.6b For Real-Time Speech

By vramkickedin | | 2 min read
A close-up object of a speech recognition module.

The Nemotron-3.5-ASR-Streaming-0.6b model is NVIDIA’s latest open speech recognition release, designed to transcribe audio in real time across 40 language-locales from a single model. It can handle both low-latency streaming and high-volume batch work, with built-in punctuation and capitalization. Users can adjust the audio chunk size at runtime — from 80 milliseconds up to 1.12 seconds — to balance speed and accuracy for different needs.

NVIDIA built this 600-million-parameter system on a cache-aware FastConformer-RNNT architecture. The company trained it on a massive mixture of public and proprietary speech data, covering 19 transcription-ready locales right out of the box. An automatic language detection mode lets a single deployment handle mixed-language traffic without a separate identification step.

Streaming efficiency and runtime control

Key capabilities
  • One model transcribes 40 language-locales.
  • Cache-aware design avoids redundant computations.
  • Configurable chunk sizes from 80ms to 1.12s.
  • Punctuation and capitalization included in output.
  • Automatic language detection with tagging option.
  • Ready for commercial use and fine-tuning.

The tool is a strong fit for teams who want to run multilingual transcription on their own hardware rather than calling cloud services. Its streaming efficiency means a single GPU can sustain hundreds of parallel real-time streams, cutting the cost per transcription. Privacy-conscious settings also benefit, since all audio stays local while still delivering competitive accuracy.

Performance notes and how it compares

At the lowest-latency 80-millisecond setting, this model sustains roughly 17 times more concurrent streams than a larger 1.1-billion-parameter buffered ASR system on the same NVIDIA H100 GPU. Word error rates stay low across the 19 transcription-ready languages, with Spanish and Italian results under 5% even at 80ms, though eight adaptation-ready locales require fine-tuning to reach production quality. Developers can follow the NeMo toolkit examples to start streaming inference or fine-tuning immediately.

"By leveraging a state-of-the-art Cache-Aware FastConformer-RNNT architecture, the model eliminates redundant overlapping computations common in traditional "buffered" streaming." — Source: Hugging Face

Project Links