Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Speech

Three surfaces share a local neural TTS engine (Kokoro-82M, ~100 MB, WAV/PCM). Network is used only for the first model download (and for optional remote TTS backends when configured):

  • Spoken replies: the assistant’s streaming output is vocalized through the speakers as it arrives.
  • Voice input: hold Space to talk; speech-to-text feeds the composer.
  • Speech synthesis: the tts agent tool and the veyyon say CLI turn text into audio files or playback.

Setup

veyyon setup speech

Installs audio dependencies and downloads the local TTS model into the tiny-models cache.

veyyon say

veyyon say "hello world"                 # play through the speakers
veyyon say --file notes.md               # speak a file
git log -1 --format=%s | veyyon say      # speak piped stdin
veyyon say "hello" --out hello.wav       # write a WAV instead of playing
veyyon say --voices                      # list models and voices
veyyon say "hello" --voice bm_fable      # pick a voice for this run

Long input is segmented into sentence-sized chunks and streamed gaplessly, so arbitrarily long text works. Error paths exit non-zero.

Spoken replies

Enable with the speech.enabled setting (Settings → Providers → Services). Related settings:

SettingWhat it does
speech.enabledSpeak the assistant’s output aloud as it streams.
speech.modeWhat to speak: all (messages + thinking), assistant (messages only), or yield (final message only).
speech.voiceKokoro voice used for spoken replies.
speech.enhancedRewrite output into natural spoken prose with the tiny model before synthesis (describes code, drops links/markdown).

Speech pauses automatically while you are talking (push-to-talk), so the assistant does not talk over you.

Voice input

SettingWhat it does
stt.enabledEnable microphone speech-to-text (hold Space to talk, or bind app.stt.toggle).
stt.languageRecognition language (default en).
stt.modelNameOn-device STT model.

Synthesis backend

SettingWhat it does
providers.ttsBackend for the tts tool: auto (prefer local, route .mp3 to xAI when credentials exist), local (Kokoro, WAV/PCM), or xai (Grok Voice, needs xAI credentials).
tts.localModelLocal TTS model (kokoro).
tts.localVoiceDefault Kokoro voice (see veyyon say --voices).
speechgen.enabledEnable the tts agent tool for speech-file synthesis.

Cache and worker recovery

Corrupt or incomplete TTS model caches are detected and re-downloaded. The synthesis worker restarts on failure so veyyon say and the tts tool can continue after a bad download.