OMI BYOK Transcription Setup

2026-04-26 · One-time setup. Eliminates the silent-fallback path that contributed to the $60/mo Gemini incident.

Why this matters: OMI's April 26, 2026 update added BYOK (bring-your-own-key) transcription. Audio routes through your Whisper / Deepgram / AssemblyAI key instead of OMI's default cloud. Costs land on a key you control, with quotas you set, and you can audit every API call. Kills the silent-fallback path that drained the $60/mo Gemini budget last time.

1. Pick a provider

Provider	Why	Get a key
Deepgram	Cheapest at scale (~$0.0043/min Nova-2). Best latency. Best for OMI's continuous capture pattern.	console.deepgram.com (free $200 credit)
OpenAI Whisper	Familiar. Mid-priced (~$0.006/min). Good enough quality.	platform.openai.com/api-keys
AssemblyAI	Premium. Speaker labels + sentiment built in. ~$0.012/min.	assemblyai.com/dashboard
Local Whisper (self-hosted)	Zero per-minute cost. Requires a public Whisper API endpoint. Apex GPU bidet could host this in theory.	Self-hosted only

Recommendation: Deepgram Nova-2. Cheapest, fastest, free $200 credit covers ~750 hours of audio. Worth the setup.

2. Set up the key in OMI

Open OMI app (mobile) or go to home.omi.me
Navigate to Settings → Developer / Advanced → Transcription Provider (exact path may vary by app version after the April 26 update)
Select "Bring Your Own Key" or "Custom Provider"
Pick provider (Deepgram / OpenAI / AssemblyAI)
Paste the API key
Optional: pick model variant (Deepgram → Nova-2 or Nova-3; OpenAI → whisper-1)
Save

If the BYOK setting isn't where the steps above suggest, check:

Settings → Privacy → Audio Processing
Settings → Apps → Transcription
Behind a "Developer Mode" toggle that needs to be enabled first

If the BYOK option still isn't visible after the April 26 update, the feature may be rolling out by region or may require updating the OMI app to the latest version. Settings → About → Check for Updates.

3. Verify it's actually live

After setup, record a short test clip (10-15 seconds) on the device. Then:

Open the chosen provider's dashboard (Deepgram console, OpenAI Usage page, AssemblyAI dashboard)
Confirm a request landed within the last 5 minutes
Note the cost — should be a few cents at most for a 15-second clip

If you don't see a request in the provider dashboard, OMI is still using the default transcription path. Re-check the BYOK setting or check OMI's logs (Settings → Developer → Logs).

4. Cost guardrails (do these before recording)

Set a monthly spend cap on the provider side. The Gemini-fallback incident hit because there was no provider-side cap to catch the runaway. Don't repeat that.

Deepgram: Console → Settings → Spending Caps
OpenAI: platform.openai.com → Usage → Limits
AssemblyAI: Dashboard → Billing → Spending Limit

Recommended cap: $15/month. That's ~3,500 minutes of Deepgram Nova-2, far above any realistic OMI capture volume and far below the $60 Gemini incident.

The provider will email you at 50% / 80% / 100% of cap by default.

5. What this changes downstream

OMI cloud / chat / Daily Recap features still work normally — only the transcription provider changed
TP3 ingest pipeline (the webhook chain) is unchanged — it receives the same transcript text regardless of who transcribed it
The Gemini fallback in Bidet is separate and stays disabled (per the 2026-04-22 lockout). BYOK is for OMI's transcription path, not Bidet's. Bidet now runs on the GPU faster-whisper image — local-only, no cloud transcription.

6. After setup

Tell me which provider you picked and what cap you set. I'll update reference_omi_platform.md in shared memory so future sessions know which provider is active and what the spend cap is.