Get started
Choose your voice mode
Sleeptalk is built for a conversation that can continue while an agent works. Your voice mode decides who handles that conversation and how work is handed off. It is separate from where you call: browser, terminal, or Discord.
| Mode | Who talks with you | Who does the work | Background work |
|---|---|---|---|
| Cascade | Your configured agent, through separate listening and speaking models | The same agent harness | Several Sleeptalk workers on eligible agents; three run at once by default, more queue |
| Cascade with Local speech | Your configured agent; Moonshine listens and Kokoro, Piper, or Pocket TTS speaks on the host | Your selected agent or local model | Same worker behavior as Cascade; local speech also shares the host's CPU |
| GPT Realtime | An OpenAI voice model | Your configured agent receives specialist requests | One specialist task at a time; a new specialist request cancels the previous one |
| GPT Live, browser preview | GPT Live continuously listens and speaks | Your configured agent handles specialist work | Background workers on eligible agents when enabled; otherwise one backend task at a time |
These describe Sleeptalk's integrations. Selecting an OpenAI speech model does not replace your agent, its tools, or its workspace with an OpenAI coding agent.
Your agent owns the conversation: Cascade
Choose a speech provider such as Cartesia, Deepgram, OpenRouter, or Local. Sleeptalk transcribes your words, passes them to your agent, and speaks its reply. Quick questions stay in the conversation; larger tasks can go to fresh worker sessions of the same agent while you keep talking.
This is the path for multiple background workers. It is available across browser, terminal, and Discord, subject to the selected agent's capabilities. “Just a model” has no tools or worker factory; command adapters also do not receive a Sleeptalk worker factory.
Speech stays on the host: Local
Local is a Cascade speech option, not a separate agent. Choose Local in settings to install the optional runtimes and download speech models. Streaming Moonshine Small and Kokoro are the defaults; Piper is a smaller voice option, and Pocket TTS adds reference voices and a custom-voice workflow.
No speech API key is needed, and call audio is processed on the host. Text still goes to your configured agent and its model. For a fully offline call after setup, use a same-machine call with Local speech and a local model such as Ollama. English only, CPU only. See Local speech.
An OpenAI model handles voice: Realtime or GPT Live
Select OpenAI Realtime as the speech service, then choose its Model. The shared settings category contains two implementations:
- GPT Realtime hears and speaks directly, handles conversational answers, and delegates workspace, tools, and current information to your agent. Terminal and browser calls have recorded live checks; Discord is preview. See OpenAI Realtime.
- GPT Live uses a separate continuous-audio implementation. The browser preview keeps listening and speaking while eligible agents run background workers. Otherwise it uses one backend task at a time. Terminal and Discord execution are rejected. Voice and language changes apply on the next call; standalone voice previews are disabled. See GPT Live.
Both require an OpenAI API credential with access to the selected model. Speech-provider charges and agent-model usage are separate. Do not infer latency, pricing, concurrency, or interruption behavior from the shared provider label.
Speaking over a reply and cancelling tool work are separate actions. Name the task you want stopped and check its status. Jobs currently end with the call; they are not persistent tasks that survive a hangup.
See the support matrix for the release contract.