Two Voices Are Not Yet Two Tracks
Full-duplex dialogue models learn turn-taking from recordings with one track per person, yet most conversations exist only as a mix. Turning a mix into that data needs Track 1 to be the same person from the first minute to the last.
Generative separation recovers the voices chunk by chunk, and every boundary re-decides which voice takes which slot. Over 26.7 hours of 25-minute calls, a quarter to a third of a floor-holder’s words leave their own track — and on these long calls the released alignment rule does not bring the error down.
Hear the Swap — and Hear It Disappear
Example 1 · The Slots Trade Places at a Chunk Boundary
Example 2 · A Different Call, the Same Failure
Decide Who, Then Generate
In two-stage generative separation, track identity is settled before diffusion: the generator copies the order of the coarse streams it receives. Anchoring acts at exactly that point.
Enrollment speech is averaged into one fixed anchor per person, eA and eB — who owns each track, for the whole conversation.
q = ⟨u1 − u2, eA − eB⟩. Its sign keeps or swaps the streams; its size is a calibrated confidence.
Speaker similarity 0.435 → 0.438, activity F1 0.849 → 0.851: quality holds while each voice stays on its own track.
Strongest When One Person Holds the Floor
On CANDOR, 52.2% of speech is one-sided. Re-identification must recognize a near-silent listener; routing compares each talking stream with its own anchor — so the gap widens as conversation grows more one-sided.
Here 20.8% of re-ID decisions fall within a 0.05 margin, against 0.8% in balanced chunks.
Example 3 · A Monologue, Two Ways to Track It
Example 4 · Re-ID Loses the Listener
From 0:10 one participant holds the floor while the other listens. Re-identification hands the talker’s voice to the listener’s track; anchored routing keeps both where they belong.
Another call, the same pattern: once one voice takes over, re-identification swaps the tracks. Routing stays put.
The Anchors Decide Who Goes Where
Swap the anchors and the score changes sign exactly. Whoever is enrolled first owns Track 1: the tracks follow the people you name.
Example 5 · Same Audio, Anchors Swapped
CANDOR remote call, held-out · {{ex.ctrl45.where}}One Router, Three Kinds of Conversation
Trained on remote calls alone, the router carries over to face-to-face and 8 kHz telephone conversations it has never seen. Its margin marks where a second look pays: verification runs on 20–36% of the output.