SimPOTrans: Adapting Offline Speech Translation Models to Simultaneous Interpretation via Reinforcement Learning

Simultaneous Speech Translation (SimulST) aims to begin translating while the speaker is still talking, producing output at very low latency while preserving translation quality as much as possible. In recent years, researchers have begun leveraging large language models (LLMs) for simultaneous interpretation, modeling the translation process as a multi-turn dialogue task and thereby achieving state-of-the-art performance. However, such conversational simultaneous interpretation approaches typically rely on specialized training strategies to acquire "listen-while-translating" behavior—either training with synthesized read/write trajectories, or jointly optimizing translation quality and latency. The former may not provide an optimal training signal, while the latter's two objectives compete with each other, easily leading to unstable optimization.

To address this pain point, the paper proposes SimPOTrans, a reinforcement learning (RL)-based framework that enables models originally built for offline speech translation to adaptively transform into conversational simultaneous interpretation models. Its core idea is not complicated: it requires neither synthesized read/write trajectories nor explicit latency optimization. Instead, through a simple two-stage RL post-training process, it rewards translation behaviors favorable to simultaneous interpretation scenarios when the model receives only partial speech input, thereby completing the adaptive transfer from offline to simultaneous interpretation.

Why Existing Conversational Simultaneous Interpretation Approaches Fall Short

To understand the value of SimPOTrans, one must first see the limitations of existing methods. Conversational simultaneous interpretation breaks translation into small generation steps, and the model must decide for itself "how far it has read and what it should write now." To help the model learn this rhythm, a common approach is to artificially construct read/write trajectories as supervision signals, but synthesized trajectories are not real simultaneous interpretation behavior after all—what the model learns may simply be imitating an imperfect "script." Another approach jointly optimizes translation quality and latency, but these two naturally exist in tension—to translate accurately often requires waiting for more context, while to achieve low latency one must output earlier. The two objectives pull against each other, and the training process is prone to instability.

SimPOTrans sidesteps these two problems. It relies on no synthesized read/write trajectories and does not treat latency as an explicit optimization term; instead, it directly converts "performing well in simultaneous interpretation scenarios" into a reward signal and lets reinforcement learning guide the model.

Read/write trajectories are a core concept in conversational simultaneous interpretation. During generation, the model alternates between two actions: reading more speech input, or writing out the translation already determined. A read/write trajectory is the sequence of these two actions. Ideally, the model should write high-quality translations when information is sufficient and continue reading when information is insufficient. But real simultaneous interpretation behavior is difficult to annotate at scale, so a common practice is to synthesize trajectories using heuristic rules or offline models as supervision signals. The problem is that synthesized trajectories themselves may deviate from the optimal policy, and a model imitating them is in effect learning from a suboptimal demonstration, which limits the performance ceiling.

Joint quality–latency optimization refers to simultaneously minimizing translation loss and latency metrics (such as average lagging words or computation-aware latency) during training. Because the two naturally conflict—waiting for more context improves quality but increases latency—joint optimization often requires manually balancing weights in the loss function, which easily leads to training oscillation or getting stuck in a local optimum for one objective. SimPOTrans's motivation is precisely to bypass this explicit trade-off.

Two-Stage RL Post-Training: Letting Offline Models Learn Simultaneous Interpretation on Their Own

The entire SimPOTrans pipeline can be summarized as "offline model + two-stage reinforcement learning post-training." In the first stage, the model is guided to produce translation behaviors beneficial to simultaneous interpretation under the condition of partial speech input; in the second stage, this adaptive capability is further reinforced. Throughout the process, the model needs neither an externally given read/write decision policy nor an explicit latency penalty term; instead, it spontaneously learns how to output when information is incomplete through the reward mechanism.

The ingenuity of this design lies in internalizing the decision of "when to translate" into the model itself. Traditional approaches often require an external decision policy to determine how much to read and how much to write, whereas SimPOTrans requires no such external policy at inference time—the model itself can handle streaming input.

RL post-training refers to further adjusting parameters with a reinforcement learning objective on top of an existing supervised pre-trained model. Unlike supervised fine-tuning, it does not require word-by-word labeled correct outputs; instead, it defines a reward function and lets the model maximize expected reward through sampling and policy gradients (such as REINFORCE or PPO). In SimPOTrans, the reward design targets the condition that "the model receives only partial speech input": when input is incomplete, behaviors that output reasonable translations early and avoid hallucination or inappropriate waiting receive higher rewards. In the two-stage process, the first stage lets the model adapt to partial input conditions, and the second stage further reinforces simultaneous interpretation behavior on that basis, enabling the offline model to gradually acquire streaming processing capability without manually constructed read/write decision labels.

Flexible Quality–Latency Control at Inference

At inference time, SimPOTrans provides a very practical adjustment mechanism: simply adjust the chunk size of speech input to flexibly trade off between translation quality and latency. The smaller the chunk, the earlier the model starts translating and the lower the latency, but the less context is available; the larger the chunk, the more sufficient the context and the more guaranteed the quality, but the latency rises accordingly. This "one knob to the end" design is quite deployment-friendly—the same model can run at different operating points according to scenario requirements, without having to train a separate model for each latency tier.

Chunk size is a key inference hyperparameter in streaming speech translation. Speech input is split into fixed-length chunks, and each time the model receives a chunk it decides whether to output a translation. The smaller the chunk, the earlier the model gets an opportunity to output and the lower the latency, but the less acoustic context is visible at each step and translation quality may decline; the larger the chunk, the more sufficient the context and the better the quality, but the later the first word is output. The reason SimPOTrans can control the trade-off solely by adjusting chunk size is that its RL post-training has already taught the model to make reasonable decisions when information is incomplete, so at inference time no additional read/write policy network or latency penalty term is needed, and the same set of parameters can cover multiple latency tiers.

Experimental Results: Quality–Latency Trade-offs Across Three Language Directions

The paper conducts experiments on two benchmarks, ACL 60/60 and MCIF, for three language directions (English→German, English→Chinese, English→Japanese). The results show that SimPOTrans outperforms multiple strong baselines and achieves better performance in the quality–latency trade-off. This demonstrates that adapting offline models to simultaneous interpretation through reinforcement learning is a feasible and effective path, and it does not rely on problematic elements such as synthesized trajectories or explicit latency optimization.

From a broader perspective, this work offers an inspiring idea: rather than painstakingly constructing specialized training signals for simultaneous interpretation, it is better to define "what behaviors are beneficial to simultaneous interpretation" as a reward and let the model figure it out on its own through reinforcement learning. For teams hoping to quickly transform existing offline translation models into streaming simultaneous interpretation capability, this post-training paradigm lowers the barrier and also avoids the instability brought by joint optimization.

Of course, the current results are concentrated on three language directions and two benchmarks, and the method's generalization ability to more languages and more scenarios remains to be further verified. But as far as solving the core problem of "offline-to-simultaneous-interpretation adaptation" is concerned, SimPOTrans provides a concise and effective route.

One-line Summary

SimPOTrans uses two-stage reinforcement learning post-training to adapt offline speech translation models into conversational simultaneous interpretation models.

Paragraph Summary

SimPOTrans proposes a reinforcement learning-based post-training framework that adaptively transforms offline speech translation models into conversational simultaneous interpretation models. To address the suboptimal training signals and optimization instability caused by existing conversational simultaneous interpretation methods relying on synthesized read/write trajectories or joint optimization of quality and latency, this method designs a two-stage RL process: when the model receives only partial speech input, it rewards translation behaviors beneficial to simultaneous interpretation, thereby achieving offline-to-simultaneous-interpretation transfer without synthesized trajectories or explicit latency optimization. At inference time, no external read/write decision policy is needed; the quality–latency trade-off can be flexibly controlled simply by adjusting the speech input chunk size. On the ACL 60/60 and MCIF benchmarks for the English-German, English-Chinese, and English-Japanese directions, SimPOTrans outperforms multiple strong baselines, validating the effectiveness of this paradigm.