Dictation and spoken replies

Open Settings → Voice on the desktop. Select On this computer and download a dictation model before choosing it. Local dictation processes recordings on this computer. Parakeet finds the language; Bark hears English; Whisper uses your chosen dictation language. The same verified models and selection work on supported Windows x64, Apple silicon and Linux x64 with glibc builds.

For Linux, the verified test environment is Ubuntu 22.04 x64 on X11 with PulseAudio. Linux arm64 and musl builds have no local runtime. Other distributions, Wayland-only sessions, PipeWire-only audio and physical microphone/headset combinations are not checked yet.

Dictate into a draft

Use the microphone in the composer, speak, then finish to insert the text. Cancel leaves the draft alone. Leaving the chat, losing the microphone or cancelling transcription releases capture and ends its native inference. Another window's voice work keeps running.

If the microphone is refused or missing, check the device and permissions, then try again. An audio capture failure never sends an unfinished turn. A failed or cancelled model download keeps a working model in place and removes its incomplete files.

Read replies aloud

Download the Supertonic 3 voice model and choose a local voice in Settings → Voice. Use the voice sample or Read aloud on a reply. Stop ends playback and pending native generation. A later reply reloads the verified model. An interrupted audio context also releases playback; the conversation remains intact.

Models download on request and keep their listed licenses: CC-BY-4.0 for Parakeet and Bark, MIT for Whisper, and OpenRAIL-M for Supertonic weights. The bundled speech runtime uses Apache-2.0 and ONNX Runtime uses MIT. Attribution and source links ship in the app's voice notices.