SCNet Pretrained Wrapper (scilib/scnet_pretrained.py)

Provides access to a frozen SCNet audio separation model, handling source ordering and sampling rate conversion. Reuse for multi‑source audio separation tasks.

# Source: https://github.com/beita6969/ScienceClaw | License: MIT | Copyright 2025 Peter Steinberger
# Provided via justclaw.it.com; keep this notice with the code.
"""Frozen four-source SCNet separation for FoR36.

The shipped checkpoint is the public MIMO-SCNet small 4-source model from
Sony Research.  This module deliberately keeps the model behind the same
mixture-only boundary as the Demucs wrapper: the sandbox sees only mixtures,
while the GPU worker owns the model code and checkpoint.  No task fitting or
target access happens here.

The upstream source tree (``src`` of SonyResearch/mimo-audio-separation) is read from ``$SCIENCECLAW_SCNET_ROOT``, else from
the default cache directory, else from ``<model root>/scnet_mimo_small/source``.

The model was trained at 44.1 kHz with source o

... (truncated -- full source via MCP)