Streaming ASR on this device
A Ferretry spike. Pick a model, load it, then play a bundled clip or speak. Nothing is uploaded: audio stays in this tab and models come straight from Hugging Face.
1 · Model
Use Wi-Fi: the Nemotron and Parakeet models are about 0.66 GB each. Try the small baseline first, then one Nemotron, then Parakeet.
2 · Run
“Live” feeds 4 LibriSpeech clips (37 s) at real speed, the way a microphone would. “Flat-out” feeds them as fast as the device can decode. Speak streams your microphone.
Words as they arrive
3 · Results
| Model | Run | RTF | Step p50/p95 ms | First word | Word lag p50/p95 | Final after end | WER | Heap |
|---|
RTF = decode time ÷ audio time (under 1 keeps up; under 0.5 is comfortable). Word lag = when a word appeared minus when it was spoken. Heap = WebAssembly memory the engine holds (Parakeet’s runtime does not expose it).