vh
76314624bb
Sub-second streaming TTS on Chatterbox-Turbo via adaptive buffer-ratchet chunking. First audio in ~0.5s (vs ~5s one-shot) with no quality compromise — chunk joins land on natural sentence pauses and the stream converges to one large near-full-context chunk within 2-3 joins. Works because the engine runs faster than realtime; the no-starvation guarantee is proven in a GPU-free simulation (tests/test_scheduler.py). - chatterbox_fast/scheduler.py: the adaptive-chunk scheduler (pure logic, no GPU) - chatterbox_fast/app.py: FastAPI server (POST /tts streaming, /voices, /health) - bench.py: streaming client (ground-truth TTFB + starvation check) - Self-contained Dockerfile (slim base + chatterbox-tts from PyPI) - Three public-domain LibriVox starter voices baked in (see voices/ATTRIBUTION.md) MIT licensed.
voices
Each *.wav in this directory is a predefined voice. The file stem becomes
the voice name returned by GET /voices and selectable via the voice request
field. The server clones the reference on the fly — no training, no enrollment.
A good reference clip is: 5–30 seconds, a single speaker, clean (minimal noise/music), 16 kHz or higher, mono. Match the clip's language to your text.
Add a voice by dropping a wav in here (or mounting your own directory at
CBF_VOICES_DIR); /voices re-scans on every call, so no restart is needed.
Licensing note
The voices shipped in this open-source repository are redistributable (public domain / explicitly licensed for redistribution). If you add your own voices, make sure you have the right to use — and, if you redistribute the image, to share — those clips. Don't ship voices of real people or copyrighted characters without permission.