ASTRAL-quantization (mirror)

An unmodified, file-for-file mirror of Plachta/ASTRAL-quantization, taken 2026-08-26. All credit to Plachta; GPL-3.0 as licensed upstream (full text in LICENSE).

From the original card: "This is a speech linguistic content quantizer [that] operates on Hubert-large features. It is trained with explicit ASR supervision to preserve more linguistic content while discarding more speaker traits."

Contents: the bsq32 and bsq2048 quantizers (full pytorch_model.bin + *_light.pth + configs). The light checkpoints are the content extractors of Seed-VC V2.

Mirrored for use in a local macOS music app — archival insurance against the original (a personal account) going away; the app downloads from the original while it exists.