Voice chat

Soulacy voice can use a free local speech sidecar while your agent continues to use any configured LLM provider.

Connect local speech

1. Install, configure, and start the bundled adapter with sy voice enable --recipe auto. Apple Silicon Macs use MLX Whisper. The portable recipe requires an existing whisper-cli build and starts with sy voice start --recipe whisper-kokoro --whisper-model /path/to/model.bin.

GPU acceleration is selected automatically when available: Metal/MLX and MPS on Apple Silicon, CUDA on supported NVIDIA hosts, or the GPU backend compiled into whisper.cpp. Use sy voice enable --recipe auto --accelerator cpu only when you need to force the compatibility fallback. 2. In Chat, click 🎤 and choose Connect a voice sidecar. 3. Enter its URL (normally http://127.0.0.1:8081) and optional voice name. 4. Save, restart the gateway, and return to Chat.

CLI users can do the same with:

sy voice providers
sy voice configure --sidecar-url http://127.0.0.1:8081 --voice af_heart
sy voice test

Talk to an agent

Click 🎤 to open a continuous voice session. Soulacy starts listening immediately, detects the end of your turn after a short silence, transcribes the recording, and sends that text through the current agent and LLM. When the LLM has completed its response, the sidecar speaks a short conversational version and Soulacy automatically resumes listening for the next turn. The same LLM call also returns the complete written answer for Chat, so tables, citations, and supporting detail remain available without being read aloud. Voice turns carry private request metadata; guidance is not appended to the visible user message and does not affect later typed turns.

The microphone is paused while the agent is thinking or speaking so the agent does not transcribe its own answer. A quiet repeating tone confirms that the LLM is still working after transcription; the voice mute control also silences this cue. Use the center control to finish a turn manually, the playback controls to pause or stop an answer, and End session to stop continuous listening. The transcript and original answer remain in the same chat history as typed turns; audio is not persisted.

The realtime OpenAI mode remains available for users who explicitly configure it. See Voice configuration for both modes, the sidecar HTTP contract, security limits, and troubleshooting.