We tested Remotion's TikTok captions template on three clips
Remotion's TikTok template transcribes a video with whisper.cpp and renders it as a vertical clip with word-by-word captions. We ran it as shipped on three 60-second clips from the open film Tears of Steel, timed both steps and checked the captions frame by frame.
Tested by us on Nov 6, 2026, on the clips and machine listed below. Every number on this page was measured in that run.
The result
It runs end to end and produces a vertical 1080x1920 video with word-by-word captions, but as shipped it is not a drop-in Shorts or Reels caption tool. On all three of our 60-second clips the render finished with exit code 0 in a median of 215 to 229 seconds, after a whisper.cpp transcription step that took a median of 126 to 453 seconds, on a busy laptop CPU with no GPU. Peak memory was about 5.9 GB for the render and 2.7 GB for the transcription.
Three things you would want to know before using it. The film clip sits in the top 452 rows of the 1920-row frame and the rest is white, not cropped to fill the screen. The captions sit in rows 1448 to 1588, which is inside the bottom 35% that third-party guides tell creators to keep clear of app buttons. And a caption was on screen at every one of our checks, including the stretches where the official subtitles show no line: the longest unbroken run in a gap was 16.5 seconds.
The words are fair but not better than a smaller model. Against the official subtitles the template's medium.en transcript scored a word error rate of 38.0%, 23.4% and 30.2%. Most of the gap is not misheard speech: whisper.cpp also writes sound descriptions such as "(dramatic music)" and "(wind whooshing)", and the template shows them as captions. With those labels removed the rates are 16.5%, 14.1% and 18.9%, next to 15.2%, 10.9% and 17.0% for the small.en model in our We tested faster-whisper captions on three dialogue clips.
What failed or surprised us
- Nothing crashed, but our first quality check was wrong. It compared the render with a 9:16 crop of the source, and the template does not crop, so it reported captions in the wrong places. We threw those numbers away and re-measured from the rendered frames themselves; the figures on this page are the second measurement.
- The template's code sets the video to objectFit cover on a white background, yet in our render the 16:9 picture came out 1080 pixels wide, top-aligned, with white below it. We did not find out why; we report what the frames show.
- Captions never go away. The code ends each caption at the start of the next one, or after a 1200 value that it meant as milliseconds but uses as frames, which is 40 seconds. So a caption stays on screen through silence until the next one replaces it.
- whisper.cpp captions sound effects. Our three clips contained 9, 2 and 5 such labels, including "[BLANK_AUDIO]", and each one is drawn as a caption in the same big white and green style as speech.
- The same 60-second job took 126 seconds on one clip and 453 seconds on another. The laptop was shared with other work and we did not isolate the cause; the CPU load before each run is stored with the results.
- The first transcription run is slow on a fresh install because the 1.4 GB medium.en model must be downloaded. Our timings do not include that download, which happened in an earlier, stopped attempt.
What we tested and how
- Tool: Remotion TikTok captions template version template-tiktok d11217ab15223bf74c4ea48e2b14ac9d68471781 2026-11-02, remotion 4.0.541, Remotion licence: free for individuals and companies of up to 3 people, a company licence above that.
- Settings compared: Template as shipped: whisper.cpp 1.6.0 with the medium.en model, CaptionedVideo 1080x1920 30 fps.
- Input clips: Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0. 60 s excerpts re-encoded to H.264 720p.
tos_a: 60 s starting at 20 s of the film, SHA-256945262dcbe106f1e...tos_b: 60 s starting at 180 s of the film, SHA-2564097773b615cddfa...tos_c: 60 s starting at 360 s of the film, SHA-2563d3a8cd015310349...
- Machine: AMD Ryzen 7 7735U (8 cores, 16 threads), integrated Radeon 680M, not used (CPU-only runs), 31.2 GB RAM, Windows 11 10.0.26300. The laptop was shared with other work, so timings are what a busy machine gives, not a clean benchmark.
- Runs: one cold run, then three timed runs per clip and setting. Run on 2026-11-06.
We cloned Remotion's TikTok captions template at the commit named above, installed it without running install scripts, and ran it as shipped: its own script to transcribe the clip with whisper.cpp 1.6.0 and the medium.en model, then its CaptionedVideo composition rendered with a pinned headless Chrome. Each clip got one cold run and three timed runs of each step. We compared the transcript with the film's official English subtitles for the same 60 seconds, and measured the rendered frames: where the picture sits, which rows hold caption ink, and whether a caption is on screen when the subtitles show no line.
By "silence" we mean a gap in the subtitles, not a sound measurement: a check counts as silent when the official subtitles show no line at that moment. Music and effects can still be playing then.
Measurements by clip
Step 1: transcribing with the template's script
| Clip | Word error rate vs subtitles | Words (subtitles / template) | Cold run (s) | Median run (s) | Peak memory (MB) |
|---|---|---|---|---|---|
| tos_a | 38% | 79 / 92 | 566.3 | 453.1 | 2699 |
| tos_b | 23.4% | 64 / 70 | 220.2 | 204.4 | 2627 |
| tos_c | 30.2% | 53 / 60 | 128.5 | 126.3 | 2630 |
Step 2: rendering the captioned vertical video
| Clip | Caption shown while speaking | Caption shown in silence | Longest caption run in silence (s) | Video rows (px of 1920) | Caption rows (px of 1920) | Output | Median render (s) | Peak memory (MB) |
|---|---|---|---|---|---|---|---|---|
| tos_a | 57 of 57 checks | 63 of 63 checks | 16.5 | 0 to 452 | 1448 to 1588 | h264 1080x1920 | 217.2 | 5870 |
| tos_b | 69 of 69 checks | 51 of 51 checks | 11 | 0 to 452 | 1448 to 1584 | h264 1080x1920 | 229.1 | 5919 |
| tos_c | 61 of 61 checks | 59 of 59 checks | 9.5 | 0 to 452 | 1448 to 1572 | h264 1080x1920 | 215.1 | 5848 |
Reproduce it
node sub.mjs
npx remotion render CaptionedVideo out.mp4 --browser-executable=<path to chrome-headless-shell>Clone the template, install its packages, put the clip in the public folder as sample-video.mp4, then run the template's transcription script and render the CaptionedVideo composition. These are the two steps we ran for each clip.
Method and limits
- Word error rate is measured against the official English subtitles after lower-casing and removing punctuation. Subtitles shorten speech and write numbers as words, so the figures overstate real mistakes; use them to compare, not as an exact score. The version without sound labels removes any text inside brackets or parentheses from the template's transcript.
- Frames were sampled twice a second. A caption counts as on screen when more than 150 dark pixels (darker than 100 of 255) appear in rows 1300 to 1700 of a 270x480 sample. The picture rows are the rows that are not white in most samples.
- The Remotion licence is free for individuals and companies of up to 3 people and needs a company licence above that. Check the current terms before you use it for work.
- Three 60-second clips of acted film dialogue with music, CPU only, on a laptop that was running other work. We tested the template as shipped, with no changes to its styling or grouping.
- Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, used under CC-BY 3.0.
Page checked on Nov 4, 2026.