Do Full-Duplex Speech Models Listen in Text? A Latent Transcript of the User and a Voice Gate That Closes It

Full-duplex speech models such as Moshi listen and speak at the same time, but emit text only for their own speech. Do they nevertheless represent what the user says as text? Applying logit and tuned lenses to Moshi and PersonaPlex, we find a latent transcript of the user in layers 21–30: the text head ranks the current user word among its top five for 46–48% of content words above a matched control, on read and meeting speech alike, while the output stays silent. The word becomes readable about 150 ms before it ends, predictable words are favoured, and errors share form rather than meaning with the target; a tuned lens finds a word code from layer 12. This transcript is closed by a voice gate: while the model speaks, it loses 50–77% of its readout of the user. The gate is triggered by speech-like audio in the model's own channel, including time-reversed speech and the model's own generated replies, but hardly by noise, an echo of the user, or a second talker in the user's channel, and activation patching shows that it is set up in layers 1–8. Ablating the transcript barely changes the reply, and the gate's behavioural effect is weaker and less specific than its effect on the representation: the transcript is readable but hardly used. Code, notebooks and results: https://github.com/gglenpoi/listen-in-text