Preprint reports fixes that reduce speech-model collapse in hour-long replays

Researchers report that two changes reduce conversational collapse in speech models that listen and speak simultaneously, without retraining, in real-time replays lasting up to an hour. Their preprint tests five recent models on long replayed conversations and live conversations with an AI partner; every evaluated model eventually degrades or exhausts its context. For models that retain only a sliding window of conversation, one fix preserves the first stored attention entry, whose removal is linked to gibberish and silence. The other discourages repetition while guarding against prolonged silence. The reported improvement concerns replays, rather than demonstrated hour-long live conversations.