X2-NativeCursor addresses the need for online text progress tracking in incremental-text streaming text-to-speech (TTS). The system tracks progress based on native speech tokens, avoiding the need for waveform decoding or acoustic processing during streaming. It utilizes a local matcher to estimate the current label position, converting revisable position estimates into a cursor that does not move backward. The design links spoken labels to their original-text spans.
Experimental results show a mean absolute error of 0.151 Chinese characters with an 80-ms lookahead, compared to 1.253 characters with a 320-ms lookahead for a waveform baseline. The real-time factor decreases from 0.3598 to 0.0180. These results demonstrate improved alignment accuracy and reduced latency.
The system was evaluated on Qwen3-TTS and CosyVoice2, with separate observers trained for each backbone. Code is publicly available on GitHub. This allows for adaptation to different TTS generators.
Source: https://arxiv.org/abs/2609.09677