IndexTTS 2.5 Launches: 2.28x Faster Inference and Zero-Shot Cross-Lingual Emotion Transfer

Xianbao_QIAN · x · 2026-08-12

IndexTTS 2.5 has been released, featuring broader language coverage, zero-shot emotion transfer, and a 2.28x improvement in inference speed.

According to its technical report, the core improvements include:

Semantic Codec Compression: Reduces the frame rate from 50 Hz to 25 Hz, halving the sequence length and significantly lowering training and inference costs.
Architectural Upgrade: Replaces the U-DiT backbone in the S2M module with a more efficient Zipformer architecture, achieving parameter reduction and faster mel-spectrogram generation.
Multilingual Extension: Introduces three cross-lingual strategies (boundary-aware alignment, token-level concatenation, instruction-guided generation) supporting Chinese, English, Japanese, Spanish, and Arabic. It enables robust emotion transfer without target-language emotional training data.
Reinforcement Learning Optimization: Applies Group Relative Policy Optimization (GRPO) in the post-training of the T2S module to improve pronunciation accuracy and naturalness.

More from Multimodal