TTS-PRISM-7B

<div align="center"> <h1>TTS-PRISM-7B</h1> <h3>A Perceptual Reasoning and Interpretable Speech Model for Fine-Grained Diagnosis</h3> <p align="center"> <a href="#-paper-link-placeholder"><img src="https://img.shields.io/badge/arXiv-Paper-b31b1b.svg" alt="arXiv"></a> <a href="https://github.com/xiaomi-research/tts-prism"><img src="https://img.shields.io/badge/GitHub-Project%20Page-181717?logo=github" alt="GitHub"></a> <a href="https://huggingface.co/xiaomi-research/TTS-PRISM-7B"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face-Model-goldenrod" alt="Hugging Face"></a> <a href="https://github.com/xiaomi-research/tts-prism/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache%202.0-007EC6.svg" alt="License"></a> </p> <p align="center"> ⭐ If TTS-PRISM is helpful to your research, please help star our GitHub repo. Thanks! 🤗 </p> </div> <br>

📖 Introduction

While generative text-to-speech (TTS) models approach human-level quality, monolithic metrics (like MOS) fail to diagnose fine-grained acoustic artifacts or explain perceptual collapse. To address this, we propose TTS-PRISM, a multi-dimensional diagnostic framework for Mandarin.

Powered by the MiMo-Audio backbone and fine-tuned on a targeted 200k-sample diagnostic dataset, TTS-PRISM embeds explicit scoring criteria and reasoning into an efficient end-to-end model. It not only predicts scores but also generates rationales explaining the specific acoustic flaws or expressive highlights.

🎯 The 12-Dimensional Evaluation Schema

Unlike generalist models, TTS-PRISM evaluates speech across a strictly defined 12-dimensional hierarchical taxonomy:

🏗 Architecture Overview

<div align="center"> <img src="https://raw.githubusercontent.com/xiaomi-research/tts-prism/main/arch_diagram.png" width="70%" alt="TTS-PRISM Architecture Diagram"> <p><em>Overall architecture of the TTS-PRISM framework.</em></p> </div>

📥 Model Download & Usage

This repository contains the model weights for TTS-PRISM-7B.

You can download the model weights using the huggingface-cli:

pip install huggingface-hub

# Download the Tokenizer
hf download XiaomiMiMo/MiMo-Audio-Tokenizer --local-dir ./models/MiMo-Audio-Tokenizer

# Download the TTS-PRISM-7B weights
hf download xiaomi-research/TTS-PRISM-7B --local-dir ./models/TTS-PRISM-7B

🚀 Running Inference

For the complete inference pipeline, data preparation, and 12-dimensional diagnostic scripts, please visit our official GitHub repository: 👉 xiaomi-research/tts-prism

✒️ Citation

If you find our work helpful, please cite our paper:

@article{wang2026ttsprism,
  title={TTS-PRISM: A Perceptual Reasoning and Interpretable Speech Model for Fine-Grained Diagnosis},
  author={Wang, Xi and Wang, Jie and Song, Xingchen and Song, Baijun and Xie, Jingran and Shao, Jiahe and Lin, Zijian and Wu, Di and Meng, Meng and Luan, Jian and Wu, Zhiyong},
  journal={arXiv preprint arXiv:2604.22225},
  year={2026}
}

⚖️ License

This project is licensed under the Apache License 2.0. Copyright (c) 2026 Xiaomi Corporation.