moss-voice-identity-repairs

# MOSS voice-acting v2 -- repaired takes For each voice profile, every take whose ECAPA speaker similarity to the voice's reference fell below 0.40, regenerated with that voice's identity LoRA (see `laion/moss-voice-identity-loras`) merged at scale 1.0 on top of the identical condition adapters at the identical lambdas. **Nothing here replaces anything.** The original takes are untouched and remain part of the corpus; low-similarity takes are kept deliberately, because they are useful for training without reference audio, which is a different and legitimate use. ## Layout ``` <voice>/part-<shard>-<chunk>.tar WebDataset: <key>.mp3 (160 kbps, 48 kHz) + <key>.json <voice>/annotations.parquet one row per repaired take <voice>/report.json the voice's before/after and cost report ``` ## Columns Every row carries three groups of columns, and none of them is derived from another: * `src_*` -- the corpus's own stored annotation of the **original** take, complete: every score component (`strength_raw`, `w_blend`, `z_containment`, `contained`, `dur`, `wer`, `spk_sim`, the z-terms, the multiplicative factors), so any later re-ranking needs zero regeneration. * `org_*` -- the **original** audio re-scored here through the same sensor stack as the repair (ECAPA, WavLM-tbr, Whisper-large-v3-turbo, the corpus's fast scorer). Stored WER came from Parakeet and is not comparable to Whisper WER, which is why the before side is re-measured rather than inherited. * `rep_*` -- the **repaired** take, measured identically. Plus `src_spk_emb` / `rep_spk_emb` (ECAPA, float16) and `rep_tbr_emb` (WavLM-tbr, float16), and `repair_adapter`, `repair_adapter_sha`, `repair_scale`, `repair_epoch` so every take records exactly which weights produced it. `audio_key` is `<gid>.cNNN.mp3` and is **not unique across runs**; `run_dir` is stored beside it and every join must use both.