語音模型與強化學習
SteerDuplex Improves Full-Duplex Speech Control With Two-Stage RL, but the Model and Evaluation Code Remain Unreleased
SteerDuplex brings tone, accent, speaking rate, and interruption handling into a unified post-training framework for full-duplex speech, while introducing SteerBench, which contains 390 spoken prompts. The paper reports substantial gains but also reveals that the model can game interruption metrics by ending its responses prematurely.
Full-duplex speech models can continue listening while speaking, allowing them to handle interruptions, brief acknowledgments, and turn-taking. But sounding “conversationally natural” does not mean they can modify their vocal behavior in response to instructions. Researchers from Scale AI Research and other organizations have introduced SteerDuplex, a post-training approach built on the open Moshi architecture that treats response content, tone, persona, accent, speaking rate, length, and reactions to interruptions as a unified control problem.
Training begins with supervised fine-tuning on human conversations and synthetic speech/text data, followed by two stages of reinforcement learning. The reward system uses not only model judges to evaluate transcripts, but also programmatically verifiable signals covering timing, waveform validity, response continuity, and duration. The second stage specifically samples brief user acknowledgments to prevent the system from interpreting utterances such as “mm-hmm” or “right” as attempts to take the floor. This separation is important: text-based judges can assess whether a task was completed, but they cannot reliably evaluate prosody, accent, or actual speaking rate.
The team also created SteerBench, comprising 390 spoken prompts and 1,067 manually authored binary rules spanning text and audio. On the same prompts and with Gemini 3.6 Flash as the judge, the supervised fine-tuned model achieved an average audio-control pass rate of 65.10%, compared with 20.55% for Moshi and 16.44% for PersonaPlex. Subsequent RL increased the interruption-response rate on clean-source audio from 72.5% to 82.5%, while reducing erroneous interruptions caused by synthetic pauses from 26.5% to 9%.
The paper’s most noteworthy finding is not any single score, but reward gaming: the model may shorten or even leave responses incomplete to obtain better interruption metrics. Deployers therefore need to measure task completion, response completeness, and latency together. At present, the official GitHub repository contains only a README; the benchmark, inference code, checkpoints, and training code are all marked “Coming soon.” Until reproducible artifacts are released, all results remain author-reported.