Server
Where the voice pipeline actually runs. Ear3 is the same product no matter which of the three substrates below you pick — same SDK, same dashboard, same webhooks. What changes is who runs the Pipecat worker that owns the audio path.
Two substrates, one Docker image
Both run the identical ghcr.io/ear3-ai/pipecat-cloud-server
image. Only the substrate — whose account it runs in — differs.
| Substrate | Who pays compute | Who does ops | When to pick |
|---|---|---|---|
| Ear3-managed (default) | Ear3 (in subscription) | Ear3 | 99% of teams. No infra work. |
| Self-host | You, your cloud | You | Audio must never leave your VPC, or > 100k interviews/mo |
Whichever substrate you’re on, <Ear3VoiceInterview> /
<Ear3Interview> are unchanged, session mint still happens on
app.ear3.ai, and interview.completed still fires from the same
webhook system. The substrate is invisible from your app’s perspective.
Ear3-managed (default)
This is what you get on any Ear3 plan without touching anything. Every session spawns a worker on our Pipecat Cloud account. You don’t see the substrate at all — it’s implicit in the SDK.
What you get without lifting a finger
- Pre-warmed workers per region — first audio frame lands before the respondent hears the first TTS syllable. Cold-start budget is measured in tens of milliseconds because we amortise a warm pool across all customers.
- Multi-region routing — respondent in Dublin gets an EU worker, respondent in SF gets us-west. WebRTC latency stays under a conversational threshold end-to-end.
- Pipeline upgrades without your redeploy — new Cartesia voice, new Deepgram model, VAD tuning — we ship to our workers and everyone gets it the next day. Self-hosters get these only when they pull a new image and redeploy.
- Provider failover — Deepgram outage → we auto-switch to fallback STT. Cartesia timeout → fallback TTS. Self-hosters build this themselves.
- No vendor secrets to manage — Deepgram, Cartesia, ElevenLabs,
OpenAI, Anthropic, Google, Daily. Seven vendor accounts + rotation.
All ours. You only hold
sk_ear3_.... - Compliance audits are ours — SOC 2, GDPR data-processing agreements. Self-hosters carry this themselves.
- End-to-end debugging when audio quality is off — we see VAD thresholds, TTS latency, buffer sizes, all telemetry. Self-hosters see only the SDK layer; everything below is your black box.
When to move off it
- Compliance requires audio never leaves your VPC — HIPAA covered entities, some fintech, gov → Self-host
- Data residency in a region we don’t offer (currently no APAC / SA) → Self-host
- Custom Python processors we don’t ship in the base image (own moderation LLM, custom STT, custom recording sink) → Self-host
- Procurement requires direct billing with Daily for the audio compute line item → rare enough that we handle it manually. Contact us and we’ll route your workspace to your own Pipecat Cloud account without you touching k8s.
The honest cost math
- < 10k interviews/month: managed is cheaper — we amortise warm pools across all customers
- 10k–50k/month: roughly parity
- 50k+/month + you already have a voice-ops SRE: self-host can reclaim ~20-30% on compute, but only if that SRE cost is already sunk
The most common mistake is “let’s self-host to control it” → three months fighting cold starts, cascading TTS failures, metrics setup, and rotating seven vendor keys. In people-hours you spent more than you saved in compute.
