Full-duplex speech models collapse around ten minutes: self-distillation plus two-resolution context cuts content and voice degradation by 71–96%
Synopsis
Full-duplex spoken dialogue models drift over long sessions, becoming slower to respond, less coherent and less consistent in voice, and long-session stability had not been targeted by dedicated training or architecture. A closed-loop simulated user quantifies this minute by minute across four public models, then privileged-context self-distillation applies lightweight LoRA adaptation and a two-resolution context (TRC) compresses older history. Across three architecturally distinct models, self-distillation mitigates degradation within the ten-minute training horizon, and adding TRC reduces spoken-content and voice degradation by 71–96% relative to baseline while extending the gains to twenty-minute sessions.
Interpretation
Long-session stability of full-duplex speech models is measured and improved as its own target, not inferred from single-turn or short exchanges. Prior evaluations covered turn-taking, interruption handling and multi-turn dialogue and noted degradation across conversational rounds, but how behavior evolves over sustained sessions of tens of minutes remained unclear. In closed-loop simulation, Moshi's response timeout rate rose from 12.7% at 7–8 minutes to 70.2% at 8–9 minutes, with cumulative terminated sessions reaching 96.7% by 10 minutes; mean conversation quality for PersonaPlex and Raon-SpeechChat fell from 4.90 and 4.62 to 1.08 and 1.03, with final-minute WER of 97.9% and 98.7%.
The degradation can be held back with a small amount of additional training, and context compression keeps the gains past the training horizon. With LoRA self-distillation alone, all three models deteriorated sharply beyond the ten-minute training range, with WER at 18–20 minutes between 88.7% and 99.6%. On PersonaPlex, MiniCPM-o 4.5 and Raon-SpeechChat, self-distillation plus TRC reduced spoken-content and voice degradation by 71–96% relative to baseline; PersonaPlex retained a conversation quality of 3.53 at 18–20 minutes, Raon-SpeechChat reached 3.29 and 44.3% WER, and MiniCPM-o 4.5 reached 4.64 and 5.8%.
The recipe distills reliable short-session behavior into long sessions and keeps earlier dialogue history in compressed form inside attention. A teacher model generates dialogue in short segments, each prefixed with a summary of the preceding conversation, and the student reproduces the teacher's predictions under teacher forcing on the concatenated trajectory. Training constrains KL divergence separately for content and control tokens and matches the hidden representations fed to the audio decoder; TRC splits a session into tiles, keeps raw tokens only inside a local window, retains one gist token per departed tile, and trains only that embedding plus LoRA adapters on Q/K/V/O.
Perspective
Evaluation covers four public models, Moshi, PersonaPlex, MiniCPM-o 4.5 and Raon-SpeechChat, with 90 simulated English conversations spanning three personas and ten everyday topics, each targeting 600 s; adaptation and the twenty-minute extension are run on PersonaPlex, MiniCPM-o 4.5 and Raon-SpeechChat, and Moshi is not separately adapted because it shares PersonaPlex's backbone. The simulated user always waits for the model to finish before responding, yields the floor, and does not generate backchannels or interruptions.
The simulator does not initiate interruptions or backchannels, so long-session stability under those user behaviors is untested. Memory recall under high compression remains limited: on the FictionalQA single-needle task TRC generally falls short of the base model's short-term recall and only becomes comparatively advantageous as sessions lengthen, and recall depends on the compression level. The limited adaptation data may have affected the diversity of the models' native full-duplex behaviors.