Ultra Video Editor 1.4.6 加入口述影像功能

A while ago I wrote about Ultra Video Editor, the free, open-source video editor I've been building. This post is about its newest feature, and the one I'm proudest of: built-in audio description.

The problem nobody fixes

Almost no video on the internet has audio description: the extra narration, like on Netflix, that tells you what is happening on screen. "She opens the door. A man in a grey coat is waiting outside." For blind and visually impaired viewers, that narration is the difference between following a story and listening to music with random sounds in it.

It's rare because it's expensive and slow. Someone has to watch the video, write a script, find the pauses, record a voice and mix it in. Small creators simply don't do it.

My wife writes children's songs, and we make music videos for them. I wanted every child to be able to follow those videos, not only hear the songs. So I built audio description into the editor.

What it does

In Ultra Video Editor 1.4.6, you press Ctrl+Alt+A, choose a video (the finished render, a clip from the timeline, or any file), and press one button.

1. Gemini watches the whole video, with its sound. Not frame by frame: the whole thing. It sees movement and what happens across shots, keeps characters consistent, reads on-screen text, and hears the singing or dialogue, so it places each description into a pause. In a song that is sung almost all the way through, it speaks briefly over the singing, so the picture is never left undescribed for long.
2. You get a script you can edit. Plain text, one line per description: time | text. Change anything you like.
3. A natural voice reads it. Microsoft's neural voices are free and need no key; Google's Chirp 3 HD voices work with a Google Cloud key.
4. The editor mixes it into the video. The music is lowered under each description. You also get a .txt for the YouTube description and a .vtt description track.

Long videos, films and series, are described in 10-minute parts, and each part knows what was described before it, so names stay the same.

Here's the beginning of the real script it wrote, in Serbian, for one of our songs ("A Song About Walking"), translated for this post:

00:01 | Title: "Iskra – A Song About Walking".
00:03 | A smiling boy.
00:06 | A playground from above, then soap bubbles against a blue sky.
00:14 | A girl in a ball pit, then a girl running through a fountain.
00:20 | The wind turns the pages of a book on the sand; people walk across a courtyard.
00:25 | A baby sleeps, then the sun lights up the grass.
00:31 | A father plays with his daughter in the park.
00:46 | In winter, a father hugs his little daughter, then the mother carries her on her back.

Thirty descriptions for a three-minute song, in natural Serbian, ending with the closing credits read out loud.

It didn't start this good

The road was bumpy, and I think that part is worth sharing.

The first version only described the beginning and the end of the video. The next one described every shot, but cut sentences to fit: "A girl walks through the forest with a." Then I switched to Serbian, and a small local vision model gave me gems like "in the sky there are bubbles", with invented words and grammar from the wrong language.

The fix was to stop describing pictures one at a time and let a model watch the whole video with its sound, the way a human describer works. The difference wasn't small. It was a different product.

The offline option is still there for videos that should never leave your computer: it now uses Qwen3-VL through Ollama, and runs even on a 6 GB laptop GPU.

Works with a screen reader, start to finish

I use a screen reader every day, so everything in the editor is built to work with JAWS and NVDA as well as with a mouse. The audio description window says what it's doing ("Sending the video: 50 percent", "Gemini is describing part 3 of 12"), the focus always lands on the next step, and a short sound plays when the work is done.

Try it, and tell me what breaks

Ultra Video Editor is free and open source (GPLv3), in English, Serbian and German:
👉 https://github.com/demirajvazi10-max/Ultra-Creative-suite

A Gemini key is free from Google AI Studio, and the Flash models have a free tier. A friend is already testing it on films. I'd love to hear how it does with your videos, especially in other languages.

If you'd like to support the project, you can do it on GitHub Sponsors.

Thanks for reading, and if you make videos, describe them. Someone is listening.