---
license: mit
tags:
  - rvc
  - voice-conversion
  - aiden
  - paion
language:
  - en
pipeline_tag: audio-to-audio
---

# AIDEN — Acoustic Intonation Decoder + Expressive Neural

Single-speaker RVC v2 model trained on Jaiden (17yo male) for the Paion project.

## Architecture

- Base: Applio 3.6.2, RVC v2, HiFi-GAN vocoder, ContentVec embedder
- Pretrain: TITAN Medium 32k (`blaise-tk/TITAN`)
- F0: rmvpe
- Sample rate: 32kHz mono

## Training

- Dataset: 537 clips, ~45 min total, 15 emotional categories
- Hardware: RTX PRO 6000 Blackwell (96GB)
- Hyperparameters: batch 16, 300 epochs, save every 25
- Wall time: ~30 min
- Best loss: g_loss 24.6 at epoch 50

## Files

- `aiden_300e_19200s.pth` — final epoch 300 checkpoint
- `aiden_300e_19200s_best_epoch.pth` — best loss checkpoint
- `aiden.index` — FAISS retrieval index for ContentVec features
- `config.json` — training config
- `model_info.json` — dataset metadata

## Inference parameters (calibrated)

- index_rate: 0.4-0.7 (higher = more Jaiden timbre, less source emotion)
- protect: 0.33
- volume_envelope (rms_mix_rate): 0.25
- f0_method: rmvpe

## Status

Trained 2026-05-06. Project subsequently pivoted to Kokoro/Chatterbox-only (without voice transfer step) — voice cloning to Jaiden's specific timbre wasn't a hard requirement. AIDEN preserved here as a reference / future option.