Deploy GPT‑SoVITS Voice‑Clone Model on Alibaba Cloud Function Compute in Minutes

GPT‑SoVITS is an open‑source generative speech model that can clone voices with high fidelity from a short audio sample. It has attracted significant community interest (over 27 K GitHub stars) and can be deployed for personal or commercial text‑to‑speech services.

Application Scenarios

Education – expressive voice interaction for language training.

Gaming – personalized character voices.

Automotive navigation – real‑time spoken directions.

Live streaming – unique digital‑human voices.

Agriculture – voice‑controlled field equipment.

Robotics – speech output for robots.

Technical Architecture Overview

Function Compute – hosts the GPT‑SoVITS inference service.

File Storage NAS – stores the pre‑trained model files.

VPC – provides a private network so Function Compute can access NAS securely.

Deployment Steps on Function Compute 3.0

Log in to the Function Compute 3.0 console.

Confirm the console version is 3.0; switch via “Experience Function Compute 3.0” if necessary.

In the left navigation, click Application .

(Optional) Click Create Application if no application exists.

Select Artificial Intelligence > Voice Clone Generation GPT‑SoVITS and click Create Immediately .

On the creation page, choose Direct Deployment , verify required permissions, ensure Function Compute and NAS are enabled, keep other settings default, and click Create Application .

In the confirmation dialog, select the billing items for Function Compute and NAS, acknowledge the disclaimer, and click Agree and Deploy .

Wait about one minute; when the status shows “Deployment Successful”, an access domain is generated. Click the domain link to start using the service.

Important Disclaimer

The cloud provider does not guarantee the legality, safety, or accuracy of third‑party models and assumes no liability for damages.

Users must comply with the model’s license, usage policies, and applicable laws, bearing responsibility for any legal issues.

Voice Training Workflow

To fine‑tune GPT‑SoVITS with a custom voice, upload a long source audio file and follow these steps:

Data preprocessing – click Data Preprocessing , upload the audio, and start preprocessing.

Text correction – click Training Voice Text Proofreading to adjust the transcript.

Model fine‑tuning – click Model Fine‑Tuning to start SoVITS and GPT training. Results are stored in NAS under GPT_weights and SoVITS_weights.

After training, go to the Voice Clone & Streaming tab, select the fine‑tuned model, and test synthesis.

Training uses default UVR5 and ASR models. Alternative models can be placed in tools/asr/models and tools/uvr5/uvr5_weights as described in the official README.

Synthesis Procedure

Select a default voice template, enter the desired text, and click Synthesize Voice .

Wait for synthesis to finish, then click Play to listen.

Further Resources

GitHub repository: https://github.com/RVC-Boss/GPT-SoVITS

Function Compute 3.0 console: https://fcnext.console.aliyun.com/

Official README: https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/cn/README.md