OmniVoice TTS Setup

OmniVoice is theLAB's text-to-speech service — a VibeVoice / Qwen3-TTS-family model running on the host apexlabs-voicebox on our private Tailscale net. It's the final stage of ADI's voice loop: an LLM produces a reply, and OmniVoice turns that text into the audio the user actually hears. The whole pipeline upstream of it lives in the fully-local voice stack.

The nice part is that OmniVoice speaks the same dialect as everything else in the fleet: an OpenAI-compatible speech endpoint. If you've called OpenAI's /v1/audio/speech before, you already know the shape — point it at our host and you get audio bytes back.

#01Hit the speech endpoint

The endpoint is POST /v1/audio/speech. On ADI the base is http://<host>:8881, so the full URL is http://<host>:8881/v1/audio/speech where <host> is apexlabs-voicebox on the tailnet. The request body is plain JSON; the response is raw audio bytes.

curl -s http://<host>:8881/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model":  "omnivoice",
    "voice":  "adi_ref",
    "input":  "Welcome back. Your garage box is online.",
    "format": "wav"
  }' \
  -o out.wav        # response body is the audio itself

Four fields carry the request. model is omnivoice; voice is adi_ref, our reference/cloned voice; input is the text to speak; and format defaults to wav. Swap format for mp3, ogg, or webm when you want a compressed stream, and swap voice if you register another reference voice later.

#02Go through ADI's tts.php instead

In production we rarely call the raw endpoint. ADI ships a PHP provider abstraction, tts.php, that fronts several backends — omnivoice, cosyvoice2, kokoro, groq, google_cloud, and browser — behind one interface. For omnivoice it defaults model to "omnivoice" and voice to the OMNIVOICE_VOICE constant ("adi_ref"), then posts {model, voice, input, ...} to {base}/v1/audio/speech — exactly the call from step 01, just centralized.

# conceptually, what tts.php does for the omnivoice provider
$payload = [
    "model"  => "omnivoice",
    "voice"  => OMNIVOICE_VOICE,   # "adi_ref"
    "input"  => $text,
    "format" => $format,           # wav by default
];
$audio = http_post_json("{$base}/v1/audio/speech", $payload);

Routing through tts.php means a single call site can fall back to another provider, or switch a user from OmniVoice to a cloud voice, without the caller changing a line.

#03Lean on the disk cache

TTS is slow and our phrasing repeats — "online", confirmations, error chimes. So tts.php caches every response on disk, keyed by sha256(provider|base|model|voice|format|language|text). A repeated phrase is a cache hit: the audio comes back instantly and the response carries an X-TTS-Cache: hit header. This is the single biggest latency win in the whole voice loop, since synthesis cost drops to a disk read for anything ADI has said before.

#Verify it works

• out.wav plays back clean speech in the adi_ref voice — not silence or a JSON error body.
• A second identical request returns near-instantly and reports X-TTS-Cache: hit.
• Through tts.php, a phrase the LLM just generated is audible end-to-end with no per-request key or auth.