Two separate things happen when you speak, and only one of them is configurable:
- Speech detection (VAD) — noticing that you're talking and when you stopped. This is always local: a tiny Silero neural network runs on your CPU for every audio frame. It ships inside the app; no setup, no network, ever.
- Transcription (STT) — turning your finished utterance into text. This is the pluggable part.
Pick the transcriber in Dashboard → Settings → Ears (STT).
The three backends
| Backend | Privacy | Speed | Accuracy | Needs |
|---|---|---|---|---|
| Local Whisper (default) | ✅ Fully local | ~1–2s per utterance | Good | Nothing — downloads on first launch, with a progress bar |
| Groq Whisper | ☁️ Cloud | ~0.3s per utterance | Best (whisper-large-v3) | A Groq key (reuses your LLM key) |
| Vosk | ✅ Fully local | Instant, streams as you speak | Weakest — not recommended | Nothing — model downloads on first use |
Local Whisper
The private default: faster-whisper running on your CPU. Choose the model
size in settings — tiny (fastest), base (the balanced default), or
small (most accurate, slower). The 1–2 seconds it takes to transcribe is
the biggest chunk of her response delay on this backend.
Groq Whisper
The same key that powers her thinking also buys dramatically better ears: whisper-large-v3-turbo in the cloud, roughly five to ten times faster than local and noticeably more accurate. It also filters non-speech well — humming and background noise get discarded instead of hallucinated into words. If you're already on Groq for the LLM and are comfortable with audio going to the same provider, this is the best experience.
Vosk
Fully offline and streaming — your words appear on screen as you say them, which neither Whisper backend does. The trade-off is real: recognition quality is clearly below Whisper. It exists for the niche where offline + live partial text matters more than accuracy.
Tuning your environment
- Echo: the orb applies echo cancellation, so speakers are fine — but headphones make barge-in crisper since she never hears herself.
- Background noise: an energy gate plus the VAD ignores fans and hums. Distinct voices (TV, roommates) can still trigger listening — that's physics, not configuration.
- Getting cut off mid-thought? She closes your turn after a short silence. Keep the sentence flowing, or just keep talking — interrupting her reply is always allowed.