Local vs. Server Transcription

The single biggest privacy choice in Spellar. This page is the full breakdown — what each mode does, what leaves your device, and how to switch.

Side by side

• Local — Whisper Large model runs on your Mac via CoreML. In private mode, your audio never leaves your device. Slower (real-time-ish on M-series). Fewer languages (17). Works fully offline.
• Server — Audio streams to our pipeline (Cloudflare AI Gateway → AssemblyAI / Whisper). Higher accuracy. Faster on long recordings. 60+ languages. Requires connection.
• Bridge (Pro+) — Default to local; re-transcribe specific segments on the server only when you choose. Best of both.

What leaves your device

Local mode

• Audio: stays on disk, encrypted at rest
• Transcript: stored in our database so it can be synchronized between the devices
• Summary (AI-generated): the transcript is sent to your chosen LLM (built-in or BYOK) when you generate a summary

<aside> 🔒 If you also disable summary generation, Spellar effectively becomes a local audio recorder.

</aside>

Server mode

• Audio: streamed in chunks to our Cloudflare-fronted pipeline, then to R2 storage (encrypted at rest)
• Transcript: stored in our database so it can be synchronized between the devices
• Summary: generated by your chosen LLM

How to switch

1. Settings → Recording → Transcription Engine → pick Local, Server, or Bridge.
2. This applies to all future recordings.