gpt-sovits-onnx-custom

GPT-SoVITS V2 ONNX bundle for gpt-sovits-onnx-rs.

Fine-tuned from lj1995/GPT-SoVITS (kaoyu_v2). Performance and quality may differ from the upstream pretrained model.

This model has been fine-tuned on copyrighted videos from the internet, so commercial use is not advised.

Layout

Directory	Description	Total size
quant/	Default bundle. INT4/INT8 weight-only quantization on applicable models.	~0.9 GB
unquant/	Full-precision bundle (no quantization).	~2.8 GB

Each directory is a self-contained model bundle. Pass the directory path to --model-path in the Rust demo.

Files use the custom_* export prefix (custom_vits.onnx, custom_t2s_encoder.onnx, etc.).

Quant vs unquant

File	Quant	Unquant	Size reduction
bert.onnx	162 MB	1142 MB	85.8%
g2pW.onnx	89 MB	606 MB	85.3%
custom_t2s_fs_decoder.onnx	83 MB	304 MB	72.6%
custom_t2s_s_decoder.onnx	77 MB	301 MB	74.6%
ssl.onnx	360 MB	360 MB	—
custom_t2s_encoder.onnx	7.5 MB	7.5 MB	—
custom_vits.onnx	155 MB	155 MB	—

ssl, custom_t2s_encoder, and custom_vits are not quantized. bert, g2pW, and the T2S decoders are.

Download

Quant (recommended, smaller):

huggingface-cli download mikv39/gpt-sovits-onnx-custom quant --local-dir ./gpt-sovits-onnx-custom-quant

Unquant (full precision):

huggingface-cli download mikv39/gpt-sovits-onnx-custom unquant --local-dir ./gpt-sovits-onnx-custom-unquant

Or clone the repo and use quant/ or unquant/ directly:

git clone https://huggingface.co/mikv39/gpt-sovits-onnx-custom

Reference audio

Each bundle includes ref.wav. Suggested reference text:

格式化，可以给自家的奶带来大量的。

Usage

cargo run --release --example gpt_sovits_demo -- \
  --model-path ./gpt-sovits-onnx-custom-quant \
  --ref-text "格式化，可以给自家的奶带来大量的。" \
  --text "你好啊，这是一个测试。"

See the project README for full build and inference instructions.