Stefan Huber
← Four kinds of careful work
Tool
Coming soon

Crosstalk: A Tool for Reviewing Who Speaks When

When two people talk over each other, transcripts and speaker labels are where reviewers disagree most. Crosstalk is my own desktop app for exactly that moment: seeing every speaker turn on one timeline, so overlaps and handovers get checked instead of guessed.

The problem it looks at

A first look

Screenshot of Crosstalk: a debate video with two speaker badges showing at once, a speakers panel with turn and overlap counts, a color-coded transcript, a timeline with waveform, spectrogram and one lane per speaker, and an inspector for editing a single word's start and end time.
Early build, October 2026, on a Swedish party-leader debate with twelve speakers. At this moment two speakers talk at once, and both badges show on the video. Each speaker gets a lane under the waveform and spectrogram, the speakers panel counts turns and seconds of overlap, uncertain words are flagged in the transcript, and the inspector retimes a single word to the millisecond.
Video: “Sammanfattning av Agendas partiledardebatt 6/5-2018” by Lu (@LuddeThl) on YouTube, licensed CC BY.

What the case study will show

  1. The review problemWhere speaker labels and turn boundaries go wrong, drawn from my work reviewing and auditing audio annotation.
  2. The tool, screen by screenA short recording of Crosstalk in use on openly licensed audio, with a walk-through of each panel.
  3. The design decisionsWhat I chose to build first, what I left out, and why a reviewer would care.
  4. How I built itPlanning and specifying the app myself, and building it with AI coding assistance.

Want an early look before the case study is published? I’m happy to walk you through the current version.

Ask for a preview