liftlab × IRC · prototype

Voice Intake Assistant

model gpt-realtime-2

transport webrtc

schema IRC intake

Session

Mode Light triage · ~6 sections Extended intake · full IRC form

Settings

Model The AI model powering the conversation. Newer models may offer better comprehension and lower latency. Voice The voice character used by the AI agent when speaking to callers. Language Caller's preferred language. "Auto" lets the agent follow the caller's lead and supports code-switching between English and Spanish. Temperature fixed in GA OpenAI removed temperature from the GA Realtime API — the model runs at a fixed value. Use the system instructions below to control behavior instead.

Turn detection

Threshold 0.50 How loud speech must be to trigger detection (0 = very sensitive, 1 = only loud/clear speech). Lower values pick up more background noise.

Advanced

Prefix padding 300 ms Audio buffered before speech is detected. Prevents the first syllable from being clipped when the caller starts talking. Silence duration 500 ms How long silence must persist before the caller's turn ends. For legal intake, try 1200–2000 ms — gives callers time to think.

Idle timeout If the agent finishes speaking and the caller stays silent, send a gentle nudge after this many ms. Off by default.

System instructions default

Edits override the mode default for this call only.

Ready

00:00

sections complete 0 / 6

Transcript

Pick a mode, click Begin call, and grant microphone access to start.

Real-time session against OpenAI gpt-realtime-2 over WebRTC. Tool calls are logged to traces/ on the server at end of call. Schema follows the Stanford IRC long-form intake questionnaire.