INTP: 250,000 Synthetic Speech Intelligibility Preference Pairs

Amphion released the INTP dataset in late 2024, providing 250,000 synthetic speech preference pairs totaling over 2,000 hours of audio. The collection spans English and Chinese languages across diverse scenarios including regular speech, repeated phrases, and code-switching contexts for speech intelligibility research.

Use Cases

• Training reward models for Reinforcement Learning from Human Feedback (RLHF) in speech synthesis using preference pairs
• Optimizing intelligibility in code-switching TTS systems using English-Chinese audio samples
• Benchmarking cross-lingual synthesis quality across diverse synthetic domains

Strengths

• 250,000 preference pairs
• Over 2,000 hours of audio content
• Includes complex code-switching and cross-lingual scenarios

Limitations

• Synthetic data origin may lack the nuanced variability of natural human speech
• Restricted to English and Chinese language pairs
• Non-commercial license restricts industrial application

Provenance

Source

Amphion (Arxiv 2501.15442, 2409.00750, 2410.06885)

Collection Method

Synthetic generation using diverse TTS model integration

Freshness

Last updated July 2025; reflects recent research in synthetic speech preference modeling.