How accurate is AI transcription? We tested 28 clips
We wrote 28 scripts, had them read aloud, transcribed them and counted every wrong word. Then we added music, noise and background talk to 9 of them.
We wrote 28 scripts, had them read aloud, transcribed them and counted every wrong word. Then we added music, noise and background talk to 9 of them.
Every transcript tool says it is accurate. We wanted a number we could stand behind, so we built a test where we knew every word in advance. This post shows how we ran it, what came back, and the kinds of mistakes worth checking for before you publish a transcript.
Between October 31 and November 3, 2026 we wrote 28 short scripts in English, Spanish, Brazilian Portuguese and Mandarin. Each was a kind of video people transcribe: a cooking clip, a travel Reel, a shop's opening hours, a voice memo, a podcast segment, a lesson. A text-to-speech voice (Kokoro, through fal.ai) read each script aloud. We put the audio under a video or saved it as an audio file, and sent it to Whisper large v3 on fal.ai.
A script compared the transcript with the original word by word. It counted numbers separately, because "fifteen" written as 15 is a style choice, not a mistake. Each clip, with its script and every difference, is shown on the page it belongs to, such as the TikTok transcript and audio to text pages.
Two limits come first, because they matter more than the result:
Fourteen of the 28 clips came back with no differences at all once numbers were set aside. The longest clips held up as well as the short ones: a 90-second talk had 2 differences in 298 words.
The mistakes fell into a few groups. Knowing them tells you where to look when you check a transcript.
Words that sound like other words. "Colander" came back as "cullender" in a pasta clip. In a bike chain clip, "turn the pedals" was right, and "spin the pedals" a few seconds later came back as "petals". The same word can be right once and wrong later in the same clip.
One small word that changes the meaning. In the freelance talk, "a landing page that brings a small shop its sales" came back as "brings a small shop at sales". It reads almost normally, which is why it is easy to miss.
Names, places and foreign words. "Miradouro da Senhora do Monte" in Lisbon came back as "Muradouro de Senhora do Monte", and "Cais do Sodré" as "Cais do Cedre". The container pioneer Malcom McLean came back as "Malcolm": the model wrote the more common spelling of his name.
Spelling and spacing. British spellings came back American: crystallises, cancelling, apologising, centimetres. "Break-bulk" came back as one word, and in Spanish "arcoíris" as "arco iris".
Informal speech made formal. In Portuguese, "pra" and "pros" came back as "para" and "para os". The meaning is the same, but the speaker's casual voice is gone. In Spanish it went further: "guárdalo para el finde", save it for the weekend, came back as "para el fin de", which stops in the middle of a phrase.
Numbers as digits. "Five dollars and eighty-six cents" came back as $5.86, "one thousand dollars" as $1,000, "five percent" as 5%. In list videos, "One. Walk up to…" came back as "1. Walk up to…". For notes that is usually better. For captions you mean to match the voice exactly, it is something to change by hand.
Clean speech is the easy case, so on November 6, 2026 we made it harder. We took four of the English clips, mixed something under the voice, and transcribed the nine results with the same engine. Nothing new was recorded: the voice is the same, so any new mistake comes from what we added.
What we added, measured as loudness against the voice:
Across the nine clips, 15 of 893 words came back different. The results surprised us:
Two lessons. First, mistakes vary from run to run: the same voice gave "petals" once and "pedals" the next time, and "cullender" once and "colander" under chatter. One clean result does not prove the next one will be clean. Second, the words that broke under chatter were the same kind that were already fragile: foreign names. Background talk made a weak spot worse instead of creating new ones.
Keep the limits in mind. Our music had no singing, and a song with lyrics competes with speech much more directly. Nine clips is a small sample, and each ran once.
On clean speech the text needs a read, not a rewrite. In our clips that meant zero to three words to fix in a clip of 20 to 90 seconds. Read it against the video before you quote it or publish it, and look first at:
We have not measured songs with singing under the voice, two people speaking to the camera at once, heavy accents, or real recordings from a street or a party. We have not run any of this on the faster turbo model the site uses now. We have also not compared this model with other transcription tools, so we make no claim that it is more accurate than any of them.
When we test those, the results will go in this post, with the date. To try it on your own video, paste a link on the video transcript generator or upload a file on video to text. The first three short videos a day are free.