Shotcut's own user guide says the timing of its automatic subtitles comes from the model, and that the project does not plan to improve it. That single sentence explains most of what people report as "Shotcut speech to text not accurate" — the captions land a little early, and the odd word comes out wrong.
This is unusual. Most software documentation describes what a feature does well. Shotcut's guide has a Known Quirks list for Speech To Text that says the quiet part out loud, and reading it tells you exactly which of your problems are fixable on your end and which are not.
What the official guide actually admits
Two sentences do the heavy lifting. On timing, the user guide states:
"Subtitle items sometimes start earlier than expected. Timing is provided by the model and tool, and we lack the skills and resources to improve this."
And on wrong words:
"Expect there to be occasional errors. Like humans and non-ideal conditions, it is not perfect. We will not take action on bug reports about some piece of audio not converting to the expected text."
That second one is a policy statement as much as a technical one: transcription mistakes are not treated as bugs. So if your plan was to file a report and wait for a fix, the guide has already answered.
The reason the timing is not Shotcut's to tune is in the same section. The guide says the feature "uses AI based on OpenAI's Whisper, courtesy of the whisper.cpp project," and notes that "Speech To Text was added in version 24.10.29." Shotcut is wiring up a speech model and placing whatever timestamps come back onto a subtitle track. The cue boundaries are the model's output, not a Shotcut setting you have missed.
The fastest fix: treat the output as a first draft
Because the timing is handed to Shotcut rather than computed by it, the practical fix is editing, not reconfiguring. The guide describes where the output lands:
"The dialog creates two jobs that appear in the Jobs panel: one to export audio and another to convert to text. The results are added to the Subtitles panel as a new top-level Subtitle Track."
From there you are editing subtitle items like any hand-authored track — which means one ordering rule matters more than anything else you do. The guide is explicit about it:
"We recommend to save subtitle editing until the timeline editing is complete because moving or cutting clips on the timeline will break the subtitle synchronization."
Generate the transcript after the cut is locked. If you run Speech To Text early and then trim a clip, the drift you get afterwards is a different problem from the model's early-start quirk, and fixing one will not fix the other.
When you are ready to burn the corrected track into the picture, note the constraint the guide gives for the Subtitle Burn-In filter: it "can only be added to the Timeline Output." Adding it to an individual clip is the single most common reason people see nothing change.
Try it right here
Runs in your browser. Your file is not uploaded anywhere.
If you would rather not rebuild a whole project just to stamp a corrected caption file onto a finished video, the tool above does the burn-in step in your browser — the video and the subtitle file are read from disk locally, nothing is uploaded. It also sidesteps the export format question covered in the FAQ below.
Other causes, in order of likelihood
1. Your source clip is variable frame rate
This one is worth ruling out first, because it produces timing weirdness that looks identical to a bad transcript but affects everything in the project. The guide's section on VFR clips says:
"Symptoms like these can be caused by Variable Frame Rate (VFR) source clips. Some users report that VFR clips work in Shotcut if the VFR is subtle. Other clips, however, are completely unusable in Shotcut."
To check, open the clip and look at the frame rate in the Properties panel. The guide's rule of thumb: "If the clip frame rate does not match a standard frame rate, then it is probably VFR. If Shotcut is able to detect a VFR clip, it will add a "(variable)" tag after the frame rate. But Shotcut can not always detect VFR clips."
The fix it gives is to convert to a fixed frame rate, and step one carries a condition you should not drop: "Set the desired Settings > Video Mode in Shotcut (do not use Automatic)." Screen recordings and phone footage are the usual suspects here.
2. You are running the small bundled model
Accuracy of the words themselves is mostly a model question, and the guide is upfront that what ships is the compact option:
"Our builds include a basic model that has decent speed and accuracy but not a big size."
You are not stuck with it: "You can download a bigger and better better brain (model) in ggml format and configure it in the Speech to Text dialog, but it will be slower." (The repeated word is in the original.) The guide does not name the bundled model or give any accuracy figures, so treat third-party claims about which model Shotcut ships with as unverified — check the Speech to Text dialog in your own install.
3. On anything but Apple Silicon, this is CPU work
This explains why a strong graphics card does nothing for transcription speed, and why the bigger model hurts more than people expect:
"Currently, the only GPU our build supports is Apple Silicon. Otherwise, it is heavily multi-threaded on the CPU."
Worth knowing because Shotcut's FAQ answer on GPU usage covers drawing the interface, hardware decoding, and hardware encoding — a reader who only sees that list would reasonably assume their GPU is helping here too. On a Windows or Linux machine, it is not.
4. Version-specific timing bugs
Shotcut's release notes record two subtitle timing fixes that are worth checking your version against. Release 25.01.25 lists "Fixed Subtitles works incorrectly after a couple of minutes with non-integer frame rates," and Release 26.8.1 lists "Fixed off-by-one frame errors in subtitle bar in Timeline and Subtitle Burn In video filter."
Both are recorded as fixed in those releases, not as current behaviour — but if you are running something older and seeing drift that grows over the length of a long clip, updating is cheaper than hand-correcting a hundred cues.
How to tell it worked
Scrub to three or four spots spread across the timeline, not just the opening. The early-start quirk is roughly constant, so it shows up everywhere; drift from a VFR source or an old version gets worse the further in you go. If the offset at minute one matches the offset at minute ten, you are looking at the model's timing and hand-editing is the answer. If it grows, chase the frame rate first.
FAQ
Why do Shotcut's auto subtitles start too early? Because the timestamps come from the speech model, not from Shotcut. The user guide lists it under Known Quirks: subtitle items sometimes start earlier than expected, and the project states it lacks the skills and resources to improve this.
Can I make Shotcut's speech to text more accurate? Partly. The guide says you can download a larger model in ggml format and configure it in the Speech to Text dialog, at the cost of speed. Wrong words are otherwise expected — the guide says it will not act on bug reports about audio not converting to the expected text.
Does Shotcut's speech to text use my GPU? Only on Apple Silicon. The guide states that is currently the only GPU the build supports, and that it is otherwise heavily multi-threaded on the CPU.
Why did my subtitles drift after I re-edited the timeline? That is a separate cause. Shotcut's guide recommends finishing timeline editing before subtitle editing, because moving or cutting clips on the timeline will break subtitle synchronisation.
What format can I export the corrected subtitles as? SRT. The guide says subtitles can be imported as SRT, VTT, ASS and SSA, but that "Subtitles are exported in SRT format." If you need WebVTT for a web player, convert the exported file with SRT to VTT afterwards.
More Shotcut troubleshooting in the fix guides, including when a Shotcut export fails or gets stuck. Everything that runs locally in your browser is listed under all Clapr tools.
Quotes above are from the Shotcut User Guide (June 2026 edition) and the official release notes, as captured in September 2026. Check the current documentation before relying on version-specific details.
