walusungungulube/krogee-vc-zephyr

Krogee VC — Seed-VC fine-tuned for "Zephyr"

A fine-tuned Seed-VC voice-conversion model that converts any spoken audio into the voice of the speaker Zephyr. It is cross-lingual - the source speech does not need to be Twi. It uses the `seed-uvit-whisper-small-wavenet` architecture (22.05 kHz, Whisper-small content encoder, BigVGAN vocoder).

The target voice was fine-tuned on clean Twi (Akan) utterances from `ghananlpcommunity/ghana-twi-synthesized-speech`.

Try the demo

Convert a clip straight in your browser:

[Demo: Krogee Voice Converter](https://huggingface.co/spaces/walusungungulube/twi-zephyr-voice-converter)

The demo includes a few bundled example clips (Twi, Ga, Ewe, Swahili, Yoruba) so you can hear the cross-lingual conversion immediately.

Files

Usage

Use the checkpoint directly with Seed-VC inference:

python inference.py \
  --source <your-speech.wav> \
  --target ref_zephyr.wav \
  --output out/ \
  --diffusion-steps 25 \
  --length-adjust 1.0 \
  --inference-cfg-rate 0.7 \
  --checkpoint ft_model.pth \
  --config config_dit_mel_seed_uvit_whisper_small_wavenet.yml

Training summary

• —Dataset: ghananlpcommunity/ghana-twi-synthesized-speech (Twi, speaker Zephyr), ~10.5k valid clips
• —Steps: 2500, batch size 2, saved every 1250 steps
• —Baseline: seed-uvit-whisper-small-wavenet (zero-shot)

License

This is a derived work of Seed-VC (GPL-3.0), verbatim in LICENSE. The training data is the publicly released ghananlpcommunity/ghana-twi-synthesized-speech dataset (see its own terms); extra pretrained weights (CAM++ / OpenVoice) are auto-downloaded by the inference stack under their respective licenses.