The most accurate AI subtitle generator? We tested 8 tools on hard audio
By Ian · · 8 min read
In our October 2026 test of eight AI subtitle and caption generators on 789 words of hard audio, Captions Generator (captionsgenerator.app, the tool I make) had the lowest word error rate: 5.7%, or 45 mistakes. Kapwing was second at 6.0% (47 mistakes) and Descript third at 6.3% (50). The top three are only two to five words apart, so they are effectively tied: a re-run of the same files could swap their order.
At the other end of the table, the gap was wide. CapCut made 96 mistakes (12.2% error), which is more than twice as many as Captions Generator, and OpenSubs made 127 mistakes (16.1% error), nearly three times as many. In our earlier speed test on clean studio audio, every tool made only 0 to 2 mistakes in 168 words, so we built this harder benchmark to see where they actually break down.
Tested on October 21, 2026 in Chrome on a Mac. Default settings and free accounts for every tool.
Results: word error rate for 8 subtitle generators
We measured accuracy using word error rate: the number of wrong, missing, or extra words divided by the 789 total words in the human-made answer keys (lower is better). The four test clips lasted 73 to 85 seconds each and covered an accented conversation (166 words), a four-person meeting (218 words), speech over loud music (174 words), and an earnings call (231 words).
| Rank | Tool | Overall | Mistakes | Accented conversation | 4-person meeting | Speech over loud music | Earnings call |
|---|---|---|---|---|---|---|---|
| 1 | Captions Generator | 5.7% | 45 | 11.4% | 6.4% | 0.6% | 4.8% |
| 2 | Kapwing | 6.0% | 47 | 13.3% | 6.0% | 0.6% | 4.8% |
| 3 | Descript | 6.3% | 50 | 14.5% | 6.9% | 0.6% | 4.3% |
| 4 | VEED | 7.2% | 57 | 16.3% | 6.9% | 0.6% | 6.1% |
| 5 | Happy Scribe | 8.4% | 66 | 15.7% | 9.6% | 0.6% | 7.8% |
| 6 | AutoSubtitles | 9.3% | 73 | 16.9% | 8.7% | 1.1% | 10.4% |
| 7 | CapCut | 12.2% | 96 | 25.9% | 12.8% | 2.3% | 9.1% |
| 8 | OpenSubs | 16.1% | 127 | 44.0% | 14.2% | 1.7% | 8.7% |
No single tool won every category. Captions Generator did best on the accented conversation, Kapwing won the four-person meeting, Descript took first place on the earnings call, and five different tools tied on the music clip.
Accented conversation
This clip features two Vietnamese speakers chatting in English about food and a city called Da Lat, with fast back-and-forth turns, laughter, and repeated words (166 words total). It was the hardest clip in our test for every single tool, producing error rates between 11.4% and 44.0%.
Captions Generator made the fewest mistakes here with 19 (11.4%), followed by Kapwing with 22 (13.3%) and Descript with 24 (14.5%). Happy Scribe made 26 mistakes, VEED made 27, AutoSubtitles made 28, CapCut made 43, and OpenSubs made 73.
Proper nouns and rapid phrasing caused the most trouble. When the speakers mentioned the city "Da Lat", Captions Generator, Kapwing, Descript, and Happy Scribe spelled it right. AutoSubtitles and VEED wrote "Dalai", CapCut wrote "the leg", and OpenSubs wrote both "that life" and "the lake". OpenSubs also turned whole phrases into different words—writing "a lot of things I was like, I ate, I think, I ate this whole" when the speaker actually said "I think I was like eight, I think, eight years old"—and dropped others entirely.
Four-person meeting
Our second clip comes from a product design meeting about the corporate image on a handset, fashion, and different markets around the world (218 words total). Because it was recorded with a single microphone in the middle of the room, the audio has noticeable room echo and people talking over each other.
Kapwing led this round with 13 mistakes (6.0% error), while Captions Generator was one word behind with 14 mistakes (6.4%). Descript and VEED tied for third with 15 mistakes each (6.9%).
The other four made more mistakes: AutoSubtitles made 19 mistakes (8.7%), Happy Scribe made 21 (9.6%), CapCut made 28 (12.8%), and OpenSubs made 31 (14.2%).
Speech over loud music
For the third test, we took NASA astronaut Nick Hague's 2019 graduation message from the International Space Station—the same clean clip from our earlier speed test—and mixed in the US Navy Band playing "Anchors Aweigh" only 3 dB quieter than his voice (174 words total). Even with the band playing right behind the speech, this proved to be the easiest clip for the AI models.
Five tools tied for first place with just 1 mistake (0.6% error): Captions Generator, Kapwing, Descript, VEED, and Happy Scribe. AutoSubtitles made 2 mistakes (1.1%), OpenSubs made 3 (1.7%), and CapCut made 4 (2.3%).
The few errors that did happen landed on names and the last words. CapCut transcribed "Nick Hague" as "Nick Cave" and turned "the moon or Mars" into "the moon or tomorrow", while AutoSubtitles spelled the astronaut's name "Haig".
Earnings call
The final clip is an enterprise software company's earnings call featuring an analyst's fast question and an executive's answer about "Q4 being showtime for the enterprise sales force", pipeline, and larger deals (231 words total). It tests how well each tool handles rapid corporate jargon.
Descript took first place with 10 mistakes (4.3% error), while Captions Generator and Kapwing tied for second with 11 mistakes each (4.8%). VEED came fourth with 14 mistakes (6.1%), and Happy Scribe finished fifth with 18 (7.8%).
In the bottom half of the table, OpenSubs made 20 mistakes (8.7%), CapCut made 21 (9.1%), and AutoSubtitles trailed with 24 mistakes (10.4%).
What the mistakes looked like
We also saw two cases of made-up text: words that weren't in the audio at all. At the end of the accented conversation, Descript turned the English phrase "yeah, only the cold weather" into a complete sentence in Finnish: "Kyllä, oli kyllä hyvä". Similarly, AutoSubtitles added the word "SAM." during the instrumental music at the end of the NASA clip.
We also noticed major formatting differences, though our scoring script did not count style choices as mistakes. CapCut left out all punctuation and capital letters in its transcripts. Meanwhile, AutoSubtitles and Happy Scribe automatically tidy up repeated words—turning natural stutters like "I just, just" into "I just"—which our verbatim human answer key counts as missing words.
Which tool should you use?
- For meetings and calls: Captions Generator, Kapwing and Descript kept their overall error rates around 6% (45 to 50 mistakes in 789 words) and were within a few words of each other on the meeting and the earnings call.
- For free SRT, VTT, or TXT subtitle files: Captions Generator lets you convert video to SRT or audio to SRT on a free account for files up to 30 minutes long, up to 30 minutes a month.
- For watermark-free burned-in captions: Captions Generator exports MP4 videos with no watermark for clips up to 3 minutes on its free plan, lets you edit any word before downloading, and supports translation into 100+ languages.
- For privacy: When you use Captions Generator, only the audio leaves your device for transcription while the video stays on your computer.
- For comparing paid tiers and limits: Read our full guide to the best AI subtitle generators for a breakdown of prices and free plans across these tools.
How we tested
- The 8 tools: We tested Captions Generator, Kapwing, Descript, VEED, Happy Scribe, AutoSubtitles, CapCut, and OpenSubs on October 21, 2026, using Chrome on a Mac.
- Settings and accounts: We used free accounts and default settings for every tool. The language was detected automatically on all runs except OpenSubs, where we manually selected English on three clips because its automatic detection took several minutes per clip on our Mac (picking English can only help its score).
- Upload workflow: Each clip was uploaded as an MP4 to each tool's normal caption or transcription flow, and we copied the text straight from the editor right after automatic transcription with no manual edits. Under the hood, Captions Generator transcribes audio using ElevenLabs Scribe v2.
- Scoring and normalisation: We scored every transcript using word error rate. Before comparing, both the tool's output and the human answer key were run through the standard English text normaliser published by OpenAI with Whisper: text is lowercased, punctuation is stripped, numbers are formatted the same way, and British/American spelling differences plus fillers like "um" and "uh" are ignored.
- Answer keys and noise: We used the human answer keys exactly as published by their academic and official sources. They contain a few disputed words of their own, which count against every tool equally—pushing all error rates up slightly without changing the ranking. Because re-running the exact same file through an AI tool can change a few words, differences of 2 to 5 mistakes between tools are within normal noise.
- Get our test data: Every audio clip, every human answer key, every tool's raw transcript and our scoring script are on GitHub.
- Audio credits and licences (clips cut from their sources and re-encoded; licences: CC BY-SA 4.0, CC BY 4.0):
- Accented conversation: Edinburgh International Accents of English Corpus (EdAcc), University of Edinburgh, transcribed by professional transcribers (CC BY-SA 4.0).
- Four-person meeting: AMI Meeting Corpus, meeting ES2004b (CC BY 4.0).
- Speech over loud music: graduation message by NASA astronaut Nick Hague (public domain; answer key: NASA's own captions), mixed with "Anchors Aweigh" by the US Navy Band (public domain).
- Earnings call: Earnings-22, published by Rev (CC BY-SA 4.0).
FAQ
Which AI subtitle generator is the most accurate?
In our October 2026 test of 8 tools on 789 words of hard audio, Captions Generator (captionsgenerator.app, the tool I make) made the fewest mistakes: 45, a 5.7% word error rate. Kapwing (6.0%, 47 mistakes) and Descript (6.3%, 50 mistakes) were within five words, so the top three are effectively tied, and a re-run could swap their order.
How accurate are AI subtitles?
On clean studio audio in our earlier speed test, every tool we tested made just 0 to 2 mistakes in 168 words (around 99% accuracy). On hard audio with accents, crosstalk, room echo, and loud background music, accuracy dropped to between 83.9% and 94.3% overall (5.7% to 16.1% word error rate).
Which caption generator handled accented English best?
On our accented clip (two Vietnamese speakers talking in English), Captions Generator made the fewest mistakes: 19 in 166 words (11.4%), ahead of Kapwing (13.3%) and Descript (14.5%). It was still the hardest clip for every tool, and one clip is a small sample, so try a minute of your own speakers before you choose.
Does background music ruin AI transcription accuracy?
No, background music barely affected modern AI subtitle generators in our test, even when mixed only 3 dB quieter than the speaker's voice. Five out of eight tools—Captions Generator, Kapwing, Descript, VEED, and Happy Scribe—made just 1 mistake in 174 words (0.6% error) on our music clip.
Why did some tools score worse on verbatim transcripts?
Some tools automatically remove natural speech repetitions, which hurts their score when compared against a strict verbatim answer key. For example, AutoSubtitles and Happy Scribe tidied up repeated words like "I just, just" into "I just", which the scoring script counts as a missing word.