Under the hoodJul 2, 2026/2 min
The voice stack behind Closer.
The facts. What powers the buyer on the other end of the line.
Closer is a voice product first. The realism of the AI buyer is the product — everything else is scaffolding. Here is what runs underneath.
The stack
- TTS model
- ElevenLabs eleven_v3 — the most expressive TTS available. Supports inline audio tags like [sigh], [dry laugh], [scoff] that render as real prosody, not text.
- STT model
- ElevenLabs scribe_v2 for batch call transcripts, scribe_v2_realtime for live captions.
- Transport
- Realtime Convai agents over WebRTC. Lower latency than WebSocket, sub-second first-audio on turn.
- Voice tuning
- stability 0.45, similarity 0.8, style 0.55, speaker boost on. Tuned so hard-mode buyers actually sound annoyed.
- Turn model
- Streaming generation — audio starts playing before the sentence finishes. No waiting, no walkie-talkie feel.
Why it matters
Reps read tone before content. If the buyer sounds flat or robotic, the training breaks — reps stop treating it like a real call. The stack above is picked, not to sound impressive on paper, but to earn the first thirty seconds of every rep's attention.