ghana-vc-Kasem_xsm
Voice conversion for Kasem (xsm). An audio-to-audio model that converts speech into the Zephyr voice while preserving the language.
This checkpoint is dedicated to Kasem: it was fine-tuned only on Kasem (xsm) clips from ghanaopenai/ghana-synthetic-speech, starting from the zero-shot seed-uvit-whisper-small-wavenet base. Use it to convert audio in Kasem. For other languages — or as a cross-lingual fallback — use ghanaopenai/ghana-vc.
Use with ghana-vc
pip install git+https://github.com/GhanaNLP/ghana-vc
export HF_TOKEN=hf_...
ghana-vc convert \
--dataset <org>/<audio-dataset> \
--output <org>/<result> \
--config-name Kasem_xsm \
--num-samples 100
The library auto-selects this model when the dataset/config name matches.
Files
| File | Description |
|---|---|
ft_model.pth | Fine-tuned DiT weights (step 1,000) |
config_dit_mel_seed_uvit_whisper_small_wavenet.yml | Model / training config |
ref_zephyr.wav | Reference utterance of the Zephyr voice in Kasem |
LICENSE | Upstream Seed-VC GPL-3.0 text, retained for attribution |
Training
- Dataset:
ghanaopenai/ghana-synthetic-speechconfigKasem_xsm(Kasem) - Base:
seed-uvit-whisper-small-wavenet(zero-shot) - Weights: the step-1,000 checkpoint. Comparing checkpoints from 500 to 3,500 steps on Asante Twi, step 1,000 kept pronunciation clearest (lowest character error rate with Omnilingual ASR CTC-300M), so it is the one published.