Spaces:
Running on Zero
Running on Zero
Commit History
-
Fix BigVGAN load against current huggingface_hub, plus audio-loading drift Running
BigVGAN.from_pretrained() crashed with "missing 2 required keyword-only arguments: 'proxies' and 'resume_download'": the vendored PyTorchModelHubMixin subclass still declares both as required in _from_pretrained, but hub_mixin no longer passes either. Download config.json and bigvgan_generator.pt and construct the model directly, which is what from_pretrained does anyway and leaves the vendored file untouched. Also fixed while auditing the rest of the startup path: - torchaudio.load is deprecated in 2.8 and its mp3 support depends on backend dispatch; use librosa, already a dependency. - Keep the demucs mix on CPU and let apply_model move chunks, matching demucs.api.Separator.separate_tensor rather than pre-moving to CUDA. - Restore upstream's "None"-string handling for mel fmax so a swapped config can't silently change mel computation. - Lower default MAX_SONG_SECONDS to 120; ZeroGPU kills the call when the declared duration elapses, and free tier only has ~300s/day total.4ca16ac -
Fix missing bigvgan.meldataset vendor file
utils.py imports MAX_WAV_VALUE from .meldataset, which wasn't copied in the original vendoring pass. Ran a full import-graph audit afterward; no other local modules.* imports are missing (the two remaining unresolved imports, JDC and the cuda alias_free_activation path, are both dead code for this model config).4c0dbbe -
Replace OmniVoice TTS with Nepali song voice-swap (Seed-VC + Demucs)
Pipeline: Demucs separates vocals/instrumental, Seed-VC's F0-conditioned singing model converts the vocals to the uploaded voice reference, then converted vocals are remixed over the original instrumental. Seed-VC's modules/ and hf_utils.py are vendored unmodified (GPL-3.0) from github.com/Plachtaa/seed-vc so upstream fixes can be dropped in directly. Prior OmniVoice TTS work is preserved on the omnivoice-v1 tag and omnivoice branch.76df492 -
Fix clone error
2434d88 -
OmniVoice CPU Space
a3003cb -
initial commit
Verified signature2b8b841