| # VoiceAcousticTransfer — P2a noise_flow (2026-07-09) | |
| - `p2a_noise_flow_scaled_lufs20_40k_best_step40000.pt` — recording-condition transfer model | |
| (conditional flow matching, hidden 1024 / depth 16 / heads 16, 40k steps, LUFS-normalized | |
| raw-DAC caches). Optimizer stripped; contains model + EMA + all metadata incl. | |
| `latent_scale` (1.652), `daps_to_idx`, `style_to_idx`. | |
| Inference: Euler t 0->1, 32 steps, joint-axis CFG 2.0; inputs loudness-normalized to -20 LUFS. | |
| - `trackb_condition_analyzer_best.pt` — standalone recording-condition analyzer | |
| (contrastive encoder + a18/recipe-parameter regression + DAPS classifier heads). | |
| - `condition_gmm.joblib` — in-the-wild condition prior (GMM over analyzer embeddings). | |
| Code, evaluation harness, and usage: https://github.com/Kiberchaika/VoiceAcousticTransfer | |
| (branch `rec-con-diffusion`); run log `8jul_acoustictransfer_evolution_report_journal.md`. | |