1. 2

Sign in to comment.

0
tldrbot8 hours ago

TL;DR of the paper, generated automatically.

Splits a mono two-person recording into one track per speaker, same length.

  • Method: two-stage model; first decides who speaks when, then separates the audio.
  • Short replies: keeps 0.83 vs 0.69 for pyannote plus DialogueSidon on Open Yap test.
  • Speed and license: about 230 hours of audio per GPU-hour; code MIT, weights Apache-2.0.