Enkidu reference production

Install the model dependencies with python -m pip install -e '.[enkidu]'. For the exact versions used for the published references, create the environment from envs/enkidu.yml instead.

protect-enkidu trains universal spectral perturbations on the dataset's full speaker cohort, then writes only the references selected by a frozen manifest. It uses local ECAPA assets and an audio-only loader, avoiding training text features and their random fallbacks.

rvcbench protect-enkidu \
  --dataset-root /absolute/path/to/data/Libritts \
  --subset-manifest /absolute/path/to/reproduction/subsets/libritts16_v1/metadata.json \
  --model-directory /absolute/path/to/SpeakerRecognition \
  --device cuda:0 --seed 42 --epochs 10 \
  --output /absolute/path/to/new/protection-run

The model directory must contain hyperparams.yaml, embedding_model.ckpt, classifier.ckpt, mean_var_norm_emb.ckpt and label_encoder.ckpt. The output directory must be new; model cache files stay inside that run.

The enkidu_audio_only_cohort_v1 protocol processes mono PCM16 references at 16 kHz with one reference per batch and retains the legacy cumulative-gradient optimization. Training uses the complete filtered speaker cohort from metadata.parquet, regardless of the number of selected output pairs. The selected identities, paths and transcripts must match that cohort.

stage_manifest.json records training-reference hashes, model assets, settings, source/runtime provenance, completed training steps, current losses, elapsed time and an estimated training time remaining. Completion requires all training steps, unchanged inputs/assets and valid selected output files. Failed stages cannot be consumed by the cloning runner. Noise is saved at protected_audio/enkidu.noise, with WAVs under protected_audio/<speaker>/.

This audio-only variant changes the historical loader's RNG consumption. Historical WAV and metric equivalence therefore require a separate comparison; a successful production run does not by itself reproduce the paper's scores.