# Vevo2 Voice Conversion on Athena

This directory contains Athena's private Vevo2 voice-conversion studio.

- Code: `open-mmlab/Amphion` commit
  `26f6883110181f1dbfe95c70a7c7dbaf4de5f42a` (MIT)
- Weights: `RMSnow/Vevo2` (CC BY-NC-ND 4.0)
- Runtime: PyTorch 2.7.1 + CUDA 12.8 inherited from Athena's tested audio
  separator image; the historical Amphion torch 2.0/cu118 pins are not used.
- Storage: `/data/voice/vevo2`; removing that directory and the test image
  removes all downloaded artifacts.
- Private UI: `http://192.168.1.212:8008` through WireGuard only.
- Output: uncompressed mono WAV, 24 kHz.

The 9 September technical gate converted the official 8.6-second speech sample
through the production HTTP API in 2.342 seconds. A warm service start loaded
the model in 12.216 seconds, and peak CUDA allocation during conversion was
5605.6 MiB. The returned 24-kHz WAV was 412878 bytes and 8.6 seconds long. The
service stores named reference voices, accepts a source clip, and returns a
transient WAV download. Jobs and generated outputs are removed after delivery.
It is deliberately not exposed on the university interface.

The installed Compose project is named `mike-ai-voice`. Rebuild or recreate it
with an explicit `VOICE_GPU_UUID` so Docker keeps the service attached to the
RTX 5080. The profile controller starts and stops the existing container; it
does not rebuild it during a mode switch.

The weights are licensed CC BY-NC-ND 4.0. This deployment is for Mike's
private, non-commercial use only. Do not use or expose it as a public or
commercial voice-cloning service.