Two bugs made the flagship workflow unable to generate:
1. torchaudio >= 2.8 routes all load/save through torchcodec, which needs
FFmpeg shared libs absent on many Windows installs ("Could not load
libtorchcodec"). F5-TTS reads its reference via torchaudio.load internally,
so this crashed every generation. Add ensure_audio_io(): probes torchcodec
once and, if broken, shims torchaudio.load/save with soundfile (libsndfile,
no FFmpeg). Own I/O (temp ref, default voice, export) uses soundfile directly.
2. F5 is a cloning model and always needs reference audio, but the flagship
ships generic characters with no Voice Reference and no default voices were
installed -> "no voice reference available" on every line. default_voice()
now falls back to the reference clip bundled with f5-tts (and returns its
known transcript, skipping the ~1.6 GB Whisper download). Clear error only
if f5-tts itself isn't importable.
Workflow: replace the instructional placeholder dialogue with a short runnable
demo script; give Characters 2 and 3 distinct profiles so it shows range out of
the box. README: document the soundfile shim, bundled default voice, and the
dependency-prune / slow-start note.
Verified end-to-end (engine load -> F5 infer -> DSP -> soundfile export).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
