shangeth/Wren-TTS-0.5B-multi-expressive-demo

Wren-TTS-0.5B-multi-expressive — Gradio demo

Expressive multilingual voice-cloning text-to-speech with 23 style tags (e.g. <happy>, <sad>, <whisper>, <confused>, <sarcastic>, <sleepy>).

Fine-tuned from `shangeth/Wren-TTS-0.5B-multi` on the Expresso dataset with multilingual replay (MLS, LibriTTS-R, VCTK, Jenny, LJSpeech) to retain the 8-language voice-cloning ability of the base model.

Pick a style tag from the dropdown, enter text, choose a reference voice, and generate. The reference voice and target text language do not have to match (cross-lingual voice cloning).

• —Model: shangeth/Wren-TTS-0.5B-multi-expressive
• —Code: github.com/shangeth/wren-tts

Expected latency on the free CPU tier

For faster inference, upgrade the Space hardware to a GPU tier.

Limitations

• —Trained on limited expressive data (Expresso, ~37 h, single English domain) — style following is decent, not perfect. Long utterances may drift out of the requested style.
• —Style tags were only seen alongside English text in fine-tuning; behaviour with other languages is undefined.

Local development

# Inside hf/expressive/space/:
python fetch_samples.py         # one-time, pulls reference clips into samples/
pip install -r requirements.txt
python app.py                   # launches on http://127.0.0.1:7860

Deploying

huggingface-cli login
python hf/expressive/push_space.py                       # pushes to default Space repo
python hf/expressive/push_space.py --space_id <user>/<name> --private

License

CC-BY-NC-4.0 (inherited from Expresso fine-tune data — non-commercial use only). Upstream model, codec, and backbone carry their own licenses — see the model card.