Which Whisper Model Should You Use for Subtitles?

Published July 4, 2026

FFmpeg Commander lets you choose the AI model it uses for transcription, from tiny and fast to large and razor-accurate. The right pick depends on your computer's memory and what you are trying to do. Here is a plain-English guide, including the fix for the most common complaint: subtitles that skip parts of the video.

What the model actually changes

Every model does the same job: listen to the audio and write the subtitles. What differs between them is a three-way trade-off, accuracy versus speed versus how much memory (RAM) they need. Bigger models are more accurate but slower and hungrier for RAM. Smaller and "4-bit" models are fast and light, but they miss more.

Why subtitles sometimes skip a part, then resume

If the captions drop out for a stretch and then pick back up, that is almost always the model being unsure of itself. When its confidence in what it is hearing falls below a threshold, it writes nothing rather than guess. Smaller and 4-bit models lose confidence more easily, especially over background music, room noise, strong accents, or quiet speech. Move up to a stronger model and those gaps usually fill right back in.

If your captions are skipping, the fix is almost always a bigger model. The 4-bit models are the most likely to skip, because they trade accuracy for very low memory use.

The models at a glance

Model	Accuracy	Speed	RAM needed	Best for
large-v3	Best	Slower	~10 GB	The best possible result, on 16 GB machines and up
large-4bit	Very good	Medium	~3 GB	Near-large quality on a lighter machine (great on 8 GB)
turbo	High	Fast	~6 GB	Speed with quality close to large
medium	Good	Medium	~5 GB	A solid all-rounder
medium-4bit	Fair	Fast	~2 GB	Quick drafts on low RAM (most likely to skip)
small	Rough	Fast	~2 GB	Rough captions you will clean up
tiny	Roughest	Fastest	~1 GB	A quick test, expect mistakes

Which one should you pick?

Best possible accuracy (16 GB RAM or more): use large-v3. Fewest mistakes, fewest skips, the gold standard.
Great accuracy on an 8 GB machine: use large-4bit. This is the sweet spot for lighter Macs, it gives you large-model quality while needing only about 3 GB, so it fits where large-v3 will not.
Lots of videos, want it fast: use turbo. Near-large quality, much quicker, if your RAM can spare around 6 GB.
Quick rough captions on a light machine: use medium-4bit or small. Fine for a first pass you will edit.
Just trying it out: use tiny. Fast and light, but expect errors.

What the "4-bit" models are for

The 4-bit models exist for one reason: to run a strong model on a computer that does not have much memory. They compress the model down to a fraction of its size. You give up a little accuracy, but you get to run something far more capable than your RAM would otherwise allow. On an 8 GB machine, large-4bit is often the best you can comfortably run, and it beats the full medium and medium-4bit models for skipping.

When it is not the model's fault

No model will caption genuine silence, background music with no speech, or two people talking over each other at the same time. In those spots, leaving the subtitles blank is the correct behavior, not a bug. If a section stays empty even on large-v3, play it back and check whether there is actually clear speech to transcribe.

"Pick the biggest model your computer can comfortably hold. Accuracy comes from the model, and most skipping problems disappear the moment you move up from a small or 4-bit one."

FFmpeg Commander runs Whisper AI right on your own computer, no cloud, no subscription. Transcribe, translate, edit, and burn in subtitles, with the model that fits your machine.

Get FFmpeg Commander — $69 →

FFmpeg Commander Video Toolbox — 2026