Transcription backends
Local vs. cloud, the daily quota, accuracy expectations, and when to pick which backend.
Quilin has two transcription backends: local (on-device models) and cloud (a managed pipeline). Both count against the same daily quota. Here's how to pick.
Local vs. cloud at a glance
| Local | Cloud | |
|---|---|---|
| Where audio goes | Stays on your Mac | Streams to managed pipeline, discarded after |
| Network needed | No (model runs offline) | Yes |
| Accuracy on clean speech | Good | Great |
| Accuracy on noisy mics / accented EN / EN-ZH code-switching | OK | Best |
| Latency | ~real-time | ~real-time |
| Counts against daily quota | Yes | Yes |
Both options return audio to text. The choice is about tradeoffs — privacy and offline-ability vs. accuracy on edge cases.
When to use local
- Sensitive notes you don't want to leave the device.
- Offline drafts — the model itself runs without a network, though the daily quota cap is still tracked server-side and refreshes when you reconnect.
- Quick captures where "good enough" beats "perfect" and you just want it in your editor now.
When to use cloud
- Long meetings or interviews where transcript accuracy compounds — small errors per minute add up over 60.
- Heavy code-switching between English and Chinese mid-sentence.
- Mic quality you don't control — calls, recordings in noisy rooms, talking through AirPods in traffic.
You can change the default per recording too — the toolbar in Recorder mode has a backend toggle.
The shared daily quota
After the 14-day free trial, the free tier gives 10 minutes of dictation per day, combined across local and cloud. Pro removes the cap entirely.
Why combine them into one cap instead of giving local an unlimited path? Two reasons:
- The tier should be one coherent shape. If local were free-unlimited and cloud were quota-gated, the obvious move would be "always use local, never enable cloud" — which means the free tier never really sees what Quilin is. We'd rather the trial show the whole product, then have the daily cap be the constraint.
- The quota itself is the product decision, not a backend-cost decision. We picked 10 min/day because it's enough for a few real captures per day without being a price-gate. Bypassing it via mode-switching defeats that intent.
The quota resets at your local midnight, enforced server-side. So a long offline trip will lock out new dictation past the daily allowance until you reconnect — plan accordingly.
What about audio storage?
Cloud transcription never stores your audio. The pipeline takes the stream, returns text, drops the audio. Local transcription doesn't touch the network at all.
Recordings you save in Recorder mode live on your Mac. Quilin doesn't sync them anywhere. To export, use the share menu on any recording — you'll get the audio file, the transcript, and any AI Action outputs as separate files in the folder you pick.
See the FAQ for the full data-handling story.