🎙️ RWKVox — RWKV-7 Text-to-Speech

RWKV-7 powered text-to-speech with voice cloning.

RWKVox uses a fine-tuned RWKV-7 language model (0.4B parameters) as the backbone for speech synthesis. Given a short text and a reference voice clip, it generates natural-sounding speech that mimics the speaker's voice.

⚠️ Disclaimer: This project is intended for learning and testing purposes only. There are no plans to further develop or improve the models.

Quick Start

For full installation instructions, usage examples, CLI reference, and audio demos, see the GitHub repository:

👉 github.com/Ed-Lovera/rwkvox

Install & Run

git clone https://github.com/Ed-Lovera/rwkvox.git
cd rwkvox
pip install -e .

# Model weights (~2.8 GB) are downloaded automatically on first run
rwkvox --text "Hello, welcome to RWKVox." --voice my_voice.wav --output hello.wav

Model Files

File	Size	Description
rwkv_tts_0.4B_best.pth	1.8 GB	RWKV-7 Speech LM (fine-tuned from RWKV-7 World 0.4B)
s3gen_meanflow.safetensors	1016 MB	S3Gen decoder (CFM flow-matching + HiFiGAN vocoder)
ve.safetensors	5.5 MB	Voice Encoder (256-dim speaker embeddings)

Model Details

Attribute	Value
Architecture	RWKV-7 (24 layers, d_model=1024)
Parameters	0.4B
Audio Limit	500 tokens (11.5 seconds)
Sample Rate	24,000 Hz
Backbone	RWKV-7 World 0.4B

License

This project is released under the MIT License.

Third-Party Licenses

This project incorporates and builds upon the following open-source components:

Component	License	Usage
RWKV-7	Apache 2.0	Model architecture; fine-tuned weights derived from RWKV-7 World 0.4B (© Bo Peng)
Chatterbox	MIT	Voice Encoder and S3Gen decoder (© 2025 Resemble AI)
flash-linear-attention	MIT	RWKV-7 attention implementation

The RWKV-7 weights included in rwkv_tts_0.4B_best.pth are derived from the RWKV-7 World 0.4B base model (Apache 2.0, © Bo Peng). The Voice Encoder (ve.safetensors) and S3Gen decoder (s3gen_meanflow.safetensors) are from the Chatterbox project (MIT, © 2025 Resemble AI).

Full license texts are available in the GitHub repository.