How we test
Every tool is checked in a real browser (headless Chromium) against a reference we didn't write: ffmpeg, an independent speech recogniser, or a signal where we planted the answer. These are the results, including where a tool falls short. Last updated 2026-10-12.
| Tool | Result and how it was measured |
|---|---|
| Loudness checker | Integrated loudness of a test recording: -19.9 LUFS, the same as ffmpeg's ebur128 filter; the normalised output re-measured by ffmpeg as promised. |
| Audio sync | Found a +1250 ms offset as +1249.9 ms, through a different-sounding microphone. |
| On-device transcription | Word error rate 2.7% on a synthetic voice sample with the standard Whisper model; 26 s to transcribe 32 s including model load. |
| Clip export | A 9:16 captioned clip came out exactly 9.65 s long (audio 9.66 s), picture 38.8 dB PSNR against the source. |
| Compressor | 11.4 MB → 3.84 MB for a 5 MB target at 33.5 dB PSNR. |
| Background noise remover | Pink noise and hum at 12 dB SNR: noise in pauses down from -37.9 to -59.4 dB, voice level -0.9 dB; noise under speech reduced much less. |
| Mouth click remover | 40 clicks planted in real speech: 38 found, click energy 7.8–8 dB lower overall (pauses repaired almost completely, clicks inside loud vowels only partly); on clean speech it changed the sound by -56 dB. |
| Script to subtitles | Against a real recogniser's output (pocketsphinx) with its mistakes ("changes" for "hinges", "beer would" for "bare wood"): 85% of script words matched, every subtitle in the script's own words. |
| Video QC checker | A test file with a planted one-frame flash, a black frame, a 100 ms audio dropout and clipping: all four found at the right timecode. |
| LUT and batch LUT | GPU result against ffmpeg's lut3d on the same .cube: mean difference 1.14/255 on a photo, while the look itself moves pixels by about 30/255. |
| BPM and beats | Synthetic tracks at 85–160 BPM within 1%, beats within about 7 ms. |
| Reframe to vertical | Synthetic clip with a moving face: face found in 100% of sampled frames and kept centred. |
| Voiceover generator | An independent recogniser (pocketsphinx) understood 76–100% of the words of a generated voiceover. |
| Video to GIF | 34–40 dB PSNR on test images; a 2 MB limit met by shrinking 800 px to 460 px. |
What these numbers don't cover
Test media is synthetic or a short sample talk, the browser is open-source Chromium (the WebM export path; Chrome and Edge export MP4), and downloads from social platforms are tested against saved page data rather than the live sites, which change without notice. Real recordings vary: use the before/after controls in each tool.
Why test like this
Because the tools run on your computer, we can't watch them work for you. Comparing every result against an outside reference before anything ships is how we know they do what each page says.