Crate uvr_core 

Shared audio separation library for the CLI and desktop application.

Most applications should use one of the focused facade crates (uvr-dsp, uvr-models, uvr-vr, uvr-roformer, or uvr-runtime). This crate is the shared implementation and exposes feature-gated modules for advanced users.

The optional CPU network API is experimental until audio-pipeline validation.

Modules

audio_io

Rust audio file I/O. Decoding preserves the source rate and channels; model preprocessing owns resampling and mono-to-stereo conversion.

dsp

CPU, FP32, centered STFT with a periodic Hann window and explicit padding. Channels are processed separately; values use the reference [frequency, time] layout.

file_task

Shared file-to-file tasks for CLI and desktop callers.

model_catalog

Identities of the four supported external weight files. No weights are embedded.

model_store

External model discovery and verified downloads, independent of audio inference.

resample

Zero-phase polyphase resampling for the fixed audio baseline. Matches SciPy’s default Kaiser (beta=5) FIR design and zero boundaries.

roformer

Fixed 1296 BS-RoFormer, CPU FP32, loading the original checkpoint in Rust. The raw network keeps the reference ISTFT length; a file scheduler must pad/crop.

runtime

Explicit inference settings shared by interactive and command-line callers.

task

vr

CPU FP32 implementations of the fixed VR mask networks.

vr_dsp

Fixed VR multi-band preprocessing and reconstruction, independent of a tensor backend. Parameters and deliberate reference corrections are documented in docs/baseline.md.

weights

File identities for reproducible experiments; no checkpoint deserialization.