Ports voicebox's chunked TTS (MIT, attribution header) with two deliberate
changes: the concat half is reworked for torch tensors (matching what our
inference helpers feed the effect chain, incl. multi-channel on the last
axis), and the sample rate comes from the engine's declared rate instead
of the first chunk (fixes a latent upstream bug).
Long text (> max_chunk_chars, default 800) splits at sentence boundaries
(abbreviation/decimal-aware, bracket tags atomic, fullwidth enders via
unicode escapes for the CJK gate) -> per-chunk generation with
deterministic seed variation (seed+i) -> linear crossfade join (default
50 ms, 0 = hard cut) -> effect chain + watermark once on the joined audio.
Wired into BOTH inference paths (OmniVoice-native _run_inference and the
engine-adapter _run_backend_inference) beside the existing [pause]
stitcher; [pause] inputs keep their dedicated path. Short text is
byte-for-byte the old single-shot path; max_chunk_chars=0 disables.
New /generate form params: max_chunk_chars (>=0, default 800),
crossfade_ms (0-1000, default 50).
Tests: 15 model-free unit tests (split priorities, abbreviation/decimal/
tag guards, crossfade math incl. multichannel + clamping) + 3 stubbed-
engine endpoint tests (long text fans out with no words lost, short text
single-shot, 0 disables). Endpoint tests validated in CI — this machine
has a pre-existing local torch/Triton segfault on any main-importing test.
Spec: voicebox deep dive 1 / parity program Wave 1.2 / #346
unlimited-length item.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
