Skip to content

v0.2.0 — Edison Singing

Latest
Latest

Choose a tag to compare

View all tags
@Scottcjn Scottcjn released this 18 May 21:58
· 28 commits to main since this release
v0.2.0
78bd23b

VintageVoice v0.2.0 — Edison Singing

F5-TTS fine-tune on Edison wax-cylinder recordings (1900–1925).

  • Dataset: 569 rows, ~2.26 hours, English Edison cylinders
  • Training: 6,800 updates, 50 epochs, 2× V100 32 GB, 5h 13m wall time
  • Base: SWivid/F5-TTS

Where the weights live

What this is

A singing model, not a speech model. Mid-training audit confirmed ~60–70% of the Edison cylinder corpus is sung material (vaudeville, parlor songs, opera, lieder). Rather than fight the data, we shipped what the data is: a model that teaches any clean reference voice to perform in 1910s-era theatrical cadence over band-limited 300–3000 Hz wax-cylinder acoustic.

What it is not

For modern-sounding spoken Sophia, use base F5-TTS. v0.3.0 Edison Spoken is on the roadmap but blocked on a different upstream archive (UCSB Spoken Word, LoC sermons/lectures) — the Edison cylinders don't contain enough modern-sounding spoken material.

Quick start

pip install huggingface_hub
huggingface-cli download AutomatedJanitor/vintage-voice \
  --revision v0.2.0-edison-singing \
  --local-dir vintage-voice-edison-singing

License

  • Code (this repo): MIT
  • Weights: CC-BY-NC-4.0 (inherited from F5-TTS base)
  • Training data: Public domain (Internet Archive)

Built on sophiacore-A620I-AX (192.168.0.136). Audit pipeline + dataset prep at /mnt/18tb/vintage_voice_edison_dataset/ on-prem.

Assets 2
Loading