NVIDIA's Parakeet V2 model is one of the fastest English ASR models on the planet. The FluidAudio team ported it to the Apple Neural Engine. Neuron Flame ships it bundled, set as the default, ready to go in 60 seconds.
Free · Open source · GPL-3.0 · 100% on-device by default
| Parakeet V2 (default) | Whisper Large v3 | |
|---|---|---|
| Languages | English only | 100+ |
| Hardware | Apple Neural Engine | GPU (Metal) |
| Speed (M2) | ~7× faster than realtime | ~2× faster than realtime |
| Accuracy (English) | Within 1% WER of Whisper Large | Best in class |
| Size on disk | 474 MB | 1.6 GB |
Parakeet V2 is English-only. If you dictate in any other language, switch to a Whisper variant on the AI Models tab — same UI, same workflow. You can keep both installed and toggle between them per-language need.
Parakeet via FluidAudio runs on Apple Silicon Macs. On Intel Macs, Neuron Flame falls back to Whisper variants. The whole UX is identical; you just won't get Parakeet's speed advantage.
Parakeet was made by NVIDIA. The Apple Silicon port is by the FluidAudio team — open source on GitHub. Neuron Flame just provides the UX wrapper. Both projects deserve more attention than they currently get.
Free, ANE-optimised, default in Neuron Flame.
Download Neuron FlameFluidAudio's Parakeet V2 implementation targets the Apple Neural Engine, which is Apple Silicon only. On Intel Macs, Neuron Flame falls back to Whisper variants. Functionally identical UX, just no ANE acceleration.
For clear English speech, Parakeet V2 is within ~1% word-error-rate of Whisper Large v3 — essentially indistinguishable in practice. For very noisy audio, heavy accents, or technical jargon, Whisper Large is still the more accurate choice.
Absolutely — FluidAudio is an open-source library you can integrate into your own Swift app. Neuron Flame just packages it with a UI, recorder, dictionary, and AI enhancement layer.
It was trained primarily on English speech and works best on what it knows. NVIDIA has multilingual variants (Parakeet RNN-T XL etc.) that the FluidAudio team may port in future.