Live intelligence

Latest AI updates

1 updates
Models

Kezek: Splitting Mono Two-Person Conversations into Per-Speaker Tracks for Full-Duplex Dialogue Data

Kezek is a model that converts mono recordings of exactly two people into separate per-speaker tracks while preserving original timing, overlaps, and short replies. It uses a two-stage design: a two-voice activity model guided by a speaker fingerprint, followed by a conditioned time-domain separator whose outputs sum to the input. Evaluations show it retains more short replies, leaks less, swaps channels less often, and runs 4-5 times faster than an open pipeline baseline. Code and weights are released under permissive licences.

speaker diarizationspeech separationdialogue data