TTS Evaluation Pipeline

A unified evaluation pipeline for comparing multiple Text-to-Speech (TTS) models.

Supported Models

Model	Type	Features
GLM-TTS	glm_tts	Zero-shot voice cloning
GLM-TTS RL	glm_tts_rl	+ RL fine-tuning
Qwen3-TTS	qwen_tts	CustomVoice (preset speakers)
Qwen3-TTS VC	qwen_tts_vc	Voice cloning (Base model)
Qwen3-TTS vLLM	qwen_tts_vllm	vLLM-accelerated (Linux only)
CosyVoice	cosyvoice	Zero-shot voice cloning
CosyVoice RL	cosyvoice_rl	+ RL fine-tuning

Quick Start

# Install all environments
python setup_models.py --all

# Run evaluation with all enabled models
python main.py

# Run specific models
python main.py -m glm_tts qwen_tts cosyvoice

Output Structure

outputs/
├── latest/           # Current run results
│   ├── glm_tts/      # Model audio outputs
│   ├── qwen_tts/
│   ├── cosyvoice/
│   ├── metrics.json
│   ├── detailed_results.json
│   └── report.md
└── history/          # Archived runs
    └── 20260206_120000/

Key Metrics

• RTF (Real-Time Factor): < 1.0 means faster than real-time
• First Token Latency: Time to first audio chunk (streaming models)
• GPU Memory: Peak memory usage during synthesis

Configuration

Edit config.yaml to customize models and settings. See CLAUDE.md for detailed documentation.

Requirements

• CUDA-capable GPU
• Conda (Miniconda or Anaconda)
• 50GB+ disk space for models

Deployment

See DEPLOYMENT.md for detailed setup instructions.

---

Claude Code Deployment Prompt

Please help me deploy this TTS evaluation project. Follow the steps in DEPLOYMENT.md:

1. Check system environment (GPU, disk space, conda)
2. Install system dependencies
3. If conda is not installed, install Miniconda
4. Clone model repositories (GLM-TTS, CosyVoice)
5. Create conda environments: python setup_models.py --all
6. Download model weights (GLM-TTS and CosyVoice)
7. Verify each environment works correctly
8. Run test: python main.py -m glm_tts

If the GPU is RTX 5080/5090, install PyTorch nightly (refer to the "RTX 50 Series" section in DEPLOYMENT.md).

Report results after each step. If errors occur, try to resolve them using the Troubleshooting section in DEPLOYMENT.md first.