X2-Turn-4B — streaming ASR + turn-taking prediction

Offline single-clip demo of x-square-robot/X2-Turn-4B-0812: bilingual (中文 / English) speech recognition plus a second head that predicts, every 80 ms, whether the speaker is idle, still talking, finished, or backchannelling.

Inference follows the authors' reference implementation (voxtral_realtime.transformers.infer_asr_turn, as used by the X2-Turn turn-demo server), adapted from the live WebSocket UI to a single-clip call.

• Code: https://github.com/X-Square-Robot/X2-Turn
• Paper: https://arxiv.org/abs/2608.10878

Example audio

• examples/en_question.wav — the quickstart clip bundled with the X2-Turn repository (Apache-2.0).
• examples/zh_*.wav — four utterances from the Easy Turn test set (Apache-2.0), covering the complete / incomplete / backchannel / wait turn states.