DNS64 reference production

rvcbench denoise-dns64 enhances the references selected by a frozen manifest. It requires an explicit local DNS64 state dict and an explicit reference directory. It does not select the latest protection run, silently fall back to another model, or download weights during inference.

rvcbench denoise-dns64 \
  --dataset-root /absolute/path/to/data/Libritts \
  --subset-manifest /absolute/path/to/reproduction/subsets/libritts16_v1/metadata.json \
  --reference-directory /absolute/path/to/protection-run/protected_audio \
  --weights /absolute/path/to/dns64-a7761ff99a7d5bb6.th \
  --device cuda:0 \
  --runtime-python /absolute/path/to/dns64-worker/bin/python \
  --output /absolute/path/to/new/denoising-run

The output directory must be new. The loader constructs the upstream DNS64 architecture without pretrained downloads, loads the local state dict with strict=True, and uses evaluation mode. The upstream pretrained definitions identify DNS64 as a 16 kHz Demucs model with 64 hidden channels.

The dns64_dataset_rate_v1 recipe mirrors the historical wrapper: normalize PCM input through torchaudio, resample to the declared dataset rate (default 16 kHz), resample to the model rate if necessary, infer, optionally mix the input using --dry in [0, 1], resample back, restore the dataset-rate length, clamp and write WAV using torchaudio's default writer. It preserves all pair IDs when several pairs share a reference, enhancing that reference once.

stage_manifest.json records the clean and protected input lineage, weight hash, source hashes, runtime packages, resampling/mixing settings and output hashes. Missing or ambiguous references fail before loading the model. Invalid inference produces a failed stage manifest. The clone runner checks the known producer format's completion counts and selected output hashes before accepting denoised_audio as its reference directory. With --runtime-python, inference runs in a finite subprocess using a JSON request/result protocol. The core process validates request, weight, worker and kernel hashes, sample identities, output content, rates and frame counts. Interpreter identity retains the virtual environment entry path even when several environments share the same underlying Python binary. --timeout-seconds defaults to 600; timeout terminates the worker and marks the stage failed. worker_result.json records incremental worker progress, and worker.log retains its diagnostics. Both direct and worker execution use the same model-only inference kernel.

Isolated model environment

DNS64 requires legacy Hydra, which conflicts with the benchmark core. Use a separate Python 3.10 worker environment with no system-site packages. The pinned packages in envs/dns64-worker-py310-cu124.txt have passed pip check and real subset inference. Direct and worker outputs match on the fixed 16-pair selection. Historical DNS64 output differs by up to 5 PCM16 units, so bitwise historical equivalence remains unestablished.

The isolated model environment can be reconstructed with Python 3.10:

python3.10 -m venv /absolute/path/to/dns64-worker
/absolute/path/to/dns64-worker/bin/python -m pip install pip==24.0 typing_extensions==4.15.0
/absolute/path/to/dns64-worker/bin/python -m pip install \
  -r envs/dns64-worker-py310-cu124.txt \
  --extra-index-url https://download.pytorch.org/whl/cu124
/absolute/path/to/dns64-worker/bin/python -m pip check