Generate speech audio (WAV/MP3) from text via the user's own free TTS
stack: local LiteLLM proxy (localhost:4000) with Gemini TTS, local Piper
(Russian), Azure Neural voices.
Use when: "озвучь текст", "озвучь это", "сделай аудио из текста",
"сгенерируй речь", "сделай mp3 из текста", "проговори текст", "text to
speech", "voice this text", "read this aloud", "generate speech".
Do NOT use for: Greek audio (greek-tts skill) or speech-to-text
(voiceink-local alias / VoiceInk app).
Generate speech audio (WAV/MP3) from text via the user's own free TTS
stack: local LiteLLM proxy (localhost:4000) with Gemini TTS, local Piper
(Russian), Azure Neural voices.
Use when: "озвучь текст", "озвучь это", "сделай аудио из текста",
"сгенерируй речь", "сделай mp3 из текста", "проговори текст", "text to
speech", "voice this text", "read this aloud", "generate speech".
Do NOT use for: Greek audio (greek-tts skill) or speech-to-text
(voiceink-local alias / VoiceInk app).
TTS — text to speech via the local LiteLLM proxy
Everything goes through one endpoint: POST http://localhost:4000/v1/audio/speech
with header Authorization: Bearer sk-local-workbot (value of
LITELLM_MASTER_KEY set in litellm/docker-compose.yml). OpenAI-compatible body:
{"model": "<alias>", "input": "<text>", "voice": "<voice>"}.
local Piper on 127.0.0.1:8177, always free, single voice ru_RU-irina-medium, Russian only
Russian last-resort fallback
tts-azure-ru
Azure Neural, voice forced to ru-RU-DmitryNeural by a proxy hook
Explicit Azure voice (any locale)
tts-azure
voice-agnostic; pass a full Azure voice id like en-US-AriaNeural
Greek
—
stop: invoke the greek-tts skill instead (it wraps tts-azure with el-GR voices and dialogue handling)
Fallback: the proxy config declares tts → tts-piper → tts-azure-ru, but in
practice router fallbacks do not fire for /v1/audio/speech (verified by the
serho bot, which implements the chain client-side). So when a model fails,
retry the next alias yourself: Russian text tts → tts-piper → tts-azure-ru;
non-Russian tts → tts-pro → tts-azure (piper can't speak it).
Workflow
Health-check before generating: curl -s http://localhost:4000/health/liveliness.
If down: cd ~/projects/dotfiles/litellm && docker compose up -d.
Run the exact invocation for the chosen alias from
backends.md — per-backend curl/python commands,
output formats, voices, quirks (the Gemini response_format crash, the
conversational-text 400), and health checks.
Text longer than ~3000 chars (Gemini hard cap 4096/req): split on sentence
or paragraph boundaries, generate per chunk, concatenate with ffmpeg —
chunking recipe is in backends.md § Long text.
Want a specific delivery (cheerful, whisper, slow, accent)? Apply the
style-control patterns from prompting.md —
Gemini natural-language prefixes, the speed param, Azure SSML.
Output and delivery
tts / tts-pro return 24 kHz mono PCM WAV. Save as .wav; for an mp3
ask: ffmpeg -i out.wav -q:a 4 out.mp3.
tts-piper: always send "response_format": "mp3" explicitly → MP3;
any other value (or the field stripped by the proxy) yields WAV.
tts-azure / tts-azure-ru return MP3.
Name the file after the content (/tmp/tts_meeting_summary.mp3, not out.mp3);
inside a project/topic directory, save there instead of /tmp.
In a Telegram-bot session, ship via mcp__bot__send_document (inline player);
for a voice bubble transcode first: ffmpeg -i out.wav -c:a libopus -b:a 32k -application voip out.ogg.
In a plain CLI session, afplay out.wav or open the file.
