SimPOTrans

SimPOTrans is a reinforcement-learning-based framework that adapts a model originally built for offline speech translation into a conversational simultaneous interpretation model.

SimPOTrans does not rely on synthesized read/write trajectories, nor does it treat latency as an explicit optimization term; instead, it guides the model via reward signals to learn simultaneous interpretation behavior.

SimPOTrans adopts a two-stage RL post-training pipeline: the first stage lets the model produce behavior conducive to simultaneous interpretation under partial speech input, and the second stage further strengthens its adaptive capability.

At inference time, SimPOTrans requires no external read/write decision policy at all; the model itself can handle streaming input.

At inference, SimPOTrans only needs to adjust the speech input chunk size to achieve a flexible trade-off between translation quality and latency.

The reason SimPOTrans can control the trade-off solely by adjusting chunk size is that RL post-training has already taught the model to make reasonable decisions when information is incomplete.

Experimental results show that SimPOTrans outperforms several strong baselines, achieving better performance on the quality-latency trade-off.

Current results are concentrated on three language directions and two benchmarks, and the method's generalization to more languages and more scenarios remains to be further verified.