Voice agents your customers can talk to today.
Parley listens, thinks and answers as people speak. Start a call on this page and judge the conversation yourself.
Natural turns. It detects when you finish speaking and replies without an awkward pause.
Stays up. If the main language model fails, a backup takes over during the call.
Cost under control. Cap each call and each device, then send callers to sign up.
Built for calls that can't drop
Each behavior below is already running in the agent you can talk to above.
- Natural turn-taking
- The agent decides when you have finished speaking and starts preparing its answer early, so replies feel immediate.
- Noise enhancement
- Background noise is removed from the caller's audio before speech recognition hears it.
- Model failover
- When the main language model returns an error, a backup model continues the call.
- Silence recovery
- If the model produces no speech for a turn, the agent answers again instead of leaving the caller waiting.
- Usage limits
- Set a time cap per call and per device. When time runs out, the page shows your sign-up button.
- Your agent, your voice
- Change the name, instructions, greeting, voice and language with environment variables. No code changes.
From caller to answer
One streaming pipeline carries every call.
- Caller speaksBrowser or app microphone
- Speech recognitionDeepgram Nova-3
- Language modelGoogle Gemini, with backup
- Speech synthesisDeepgram Aura-2
- Agent answersStreamed back live
Audio travels over LiveKit. Silero voice detection and LiveKit turn detection decide when each side is speaking.