Published July 6, 2026
Whisper AI is remarkably good at transcribing and translating speech, but it has two quirks that show up in real-world content — phantom words that appear over silence or music, and subtitles that flash too fast to read during rapid-fire dialogue. FFmpeg Commander now gives you direct controls for both.
Whisper processes audio in 30-second windows. When a window contains background music, sound effects, or a quiet pause, the model does not output nothing — it invents something. These invented subtitles are called hallucinations, and they look like real text but have no matching speech in the audio.
Common examples: a single word appearing during a film's opening music, a short phrase over an action scene, or a line that repeats a few times in a row for no reason. The more complex the audio — heavy film scores, crowd noise, overlapping voices — the more likely hallucinations are to appear.
When characters talk quickly, Whisper produces many short subtitle lines in rapid succession. Each one might only stay on screen for a fraction of a second — far too short to read. You get the text, but it flashes past before your eyes can land on it. This is especially common in foreign-language films where dialogue is dense and the translated English lines are longer than the original.
Open the Transcribe (Whisper) dialog and you will find a new row just below the Performance setting:
This controls how aggressively the app removes subtitle lines that look like hallucinations. It works by tuning three internal Whisper confidence scores — the probability that a segment contains no real speech, how repetitive the output is, and the model's overall confidence in what it heard. Higher aggressiveness means fewer phantom subtitles, but at the cost of occasionally removing a real line that the model was uncertain about.
| Setting | What it does | Best for |
|---|---|---|
| Lenient | Keeps almost everything — only drops segments the model is very sure are silence | Clean studio recordings, podcasts, interviews |
| Balanced | The default — works well for most content | General use, most videos |
| Strict | Removes more — good for content with background music or noise | Films, TV shows, content with music beds |
| Very Strict | Most aggressive — removes anything the model is not confident about | Heavy background noise, overlapping audio |
Start with Strict for films. If you are transcribing or translating a feature film with a soundtrack, Strict removes the majority of hallucinations without cutting real dialogue. Only move to Very Strict if phantom subtitles are still appearing after a Strict pass.
This sets a floor on how long each subtitle line stays on screen. When a subtitle's natural duration is shorter than the setting you choose, the app extends it — up to but never overlapping the next subtitle — so the viewer actually has time to read it.
| Setting | What it does |
|---|---|
| Auto | Lets the reading-speed algorithm decide — usually fine for normal pacing |
| 1.0 s | Each subtitle stays on for at least one second |
| 1.5 s | Good starting point for fast dialogue |
| 2.0 s | Comfortable reading time for most viewers — recommended for dense translated dialogue |
| 2.5 s | For very fast speech or longer translated lines |
| 3.0 s | Maximum — use when lines are long and speech is very rapid |
For translated foreign films, 2.0 s is a good default. Translated lines are often longer than the original spoken words, so the natural Whisper timing is frequently too short. A 2-second floor gives most viewers enough time to read a typical subtitle line comfortably.
Both controls are saved when you save a preset in the Transcribe dialog. So you can set up a "Films" preset with Strict and 2.0 s and switch to it instantly the next time you transcribe a movie — without adjusting anything manually.
FFmpeg Commander runs Whisper AI locally on your own machine — no cloud, no subscription. Transcribe, translate, filter, and burn subtitles into any video with full control over the output.
Get FFmpeg Commander — $69 →Related: which Whisper model to use and how to translate foreign-language video.
FFmpeg Commander Video Toolbox — 2026