Voice & language

OpenPhn supports voice engines at call runtime. The default engine runs Gemini 3.1 Flash Live (native audio, voice Aoede); Grok Voice Think Fast 1.0 is available as an experimental challenger. The engine is resolved per-call by walking a precedence chain — per-Number override → per-Tenant default → system default — so you can opt one phone line into an experimental engine without touching the rest of your account.

Available engines

Dispatch precedence (highest priority first)

When OpenPhn dispatches a call, it picks the voice engine by walking this chain:

1. Per-Number override — numbers.voice_engine, set via PATCH /v1/numbers/{id} { "voice_engine": "grok_think_fast" }. Useful when one phone line is a test bed.
2. Per-Tenant default — tenants.voice_engine_default, set via PATCH /v1/tenants/me { "voice_engine_default": "gemini_31" } (owner only). Applies to every call placed without a Number override.
3. System default — gemini_25 until OpenPhn promotes a challenger.

If the Tenant has set voice_ab_disabled: true, the chain short-circuits to the system default — useful as a kill switch if an experimental engine regresses in production.

Choose a voice on call create

The voice_engine field on POST /v1/calls is optional. When you set it, the call uses that engine directly (bypassing the precedence chain for that one call):

{
  "to": "+14155551234",
  "objective": "...",
  "outcome_schema": { ... },
  "consent_type": "prior_express",
  "voice_id":     "puck",
  "voice_engine": "grok_think_fast",
  "language":     "en-US"
}

• voice_engine — gemini_25, gemini_31, or grok_think_fast. Omit to use the precedence chain.
• voice_id — picked from GET /v1/voices (each entry lists which engines support it).
• language — BCP-47 (en-US, es-US, fr-FR, etc.).

List voices

curl https://api.openphn.com/v1/voices \
  -H "Authorization: Bearer sk_live_..."

Each entry includes provider, gender, language, supported engines, and a preview URL.

A/B testing engines

OpenPhn runs a continuous A/B on the system default. Telemetry is captured per engine: completion rate, transfer rate, error rate, p50/p95 duration, cost per completed call, error rate, plus a Sentry tag and a PostHog property on every call_completed event.

If you're on the experimental cohort and want to opt out, set voice_ab_disabled: true on the tenant. If you want to opt INTO an experimental engine, set voice_engine_default explicitly.

OpenPhn evaluates engine promotion against the voice-engine promotion runbook — a challenger must clear four gates (volume, lift, cost, stability) over a 30-day window before it can become the system default. The next promotion review is 2026-06-22.

Failover & reliability

OpenPhn guards against a voice provider being down at call time:

• Engine health gating + connect-time failover. Before a call connects, OpenPhn checks each engine's recent health. If the chosen engine can't establish the live audio session, OpenPhn retries the next healthy engine on the same call before any audio has started — so the caller never hears a failed connection. A background probe restores an engine once it recovers.
• TTS / STT provider chains. Speech synthesis and transcription each run an ordered provider chain (configurable under Settings → Voice providers, e.g. cartesia, elevenlabs and deepgram, assemblyai). The first healthy provider wins; a provider whose recent success rate drops below threshold is skipped until a probe recovers it.