How AI audio description works
Audio description has always been slow because a person has to watch every second, write a script and record it. We automate the repetitive parts and keep you in control of the words.
1. Speech and pause detection
We transcribe the soundtrack with word-level timing. Every gap between spoken words longer than about one and a half seconds becomes a candidate slot for a description.
2. Visual analysis
A vision model reviews the frames between slots and lists what a sighted viewer would notice: on-screen text and numbers, chart values, who is speaking, actions, and scene changes. Information the narrator already says out loud is skipped.
3. Script writing
Descriptions follow common practice from established description guidelines: present tense, objective, no interpretation, most important information first, and short enough to fit the slot.
4. Standard or extended
If a description fits a pause, it plays without changing the video (standard description, WCAG 1.2.5). If the narration never pauses, as in most lectures, we pause the picture at the end of a sentence, speak the description, and resume (extended description, WCAG 1.2.7).
5. Quality checks and voice
Before export, an automatic check compares each description with its frames, removes anything the narrator already says, and flags lines that are too long for their slot. A natural-sounding voice then reads the script. You can still edit any line yourself and rebuild the files in seconds.
6. Export
- Described MP4 (standard or extended)
- Audio-only description track for YouTube's descriptive audio option or your player
- WebVTT description file (
kind="descriptions") - Editable script (DOCX and TXT)
- Compliance note listing WCAG 1.2.3, 1.2.5 and 1.2.7 coverage