Why is my transcript wrong? Causes and fixes

Speech-to-text does well with one person talking clearly and struggles with almost everything else. When a transcript comes out wrong, the cause is nearly always in the audio: music, noise, people talking over each other, an unusual name, or a language the tool didn’t expect. Find what you’re seeing in the table, then go to the fix.

Quick diagnosis

Wrong language, or gibberish

Automatic detection listens to the start of the video. If the first seconds are music, a sound effect or a line in another language, it can pick the wrong one. Then the model may write a rough translation instead of the actual words, or plain nonsense.

Fix:

1. Look at the detected language, on the line above the transcript (for example, “Spanish, detected”).
2. If it’s wrong, tap Another language and choose the right one: English, Spanish or Portuguese.
3. The new transcript replaces the old one when it arrives. Your first language change for each video is free that day.

For other languages, automatic detection is the only option. Videos that switch languages halfway through tend to come out in just one. Transcribe them anyway and fix the other parts by hand.

Names, brands and slang spelled wrong

The model writes what it hears using words it already knows. “Kaitlyn” becomes “Caitlin”, a product name splits into two ordinary words, and new slang turns into something that sounds close.

Fix: download the TXT or SRT and use find and replace in a text editor. If the word repeats, one replace fixes every line. In an SRT, change only the text and leave the time lines alone.

Song lyrics, or missing lines

Music with vocals gets transcribed along with the speaker, because the model can’t tell a singer from someone talking. Loud music can also bury quiet speech, and those lines go missing.

Fix: there’s no setting that filters music out. Delete the lyric lines by hand. If you used a repost with a soundtrack added, try the original post, which often has cleaner audio.

Empty text, or a phrase nobody said

The tool transcribes audio. It doesn’t read captions or text on the screen, so a video with only music and on-screen text has nothing to transcribe. With no speech at all, speech models sometimes output a stock phrase like “Thanks for watching” or a single stray word.

Fix: if nobody talks in the video, there’s no transcript to get. Copy the on-screen text by hand.

Two people talking at once

When people talk over each other, their words end up in the same line, and the transcript doesn’t say who is speaking.

Fix: add the names yourself. In an SRT, a common style for two speakers in one caption is a dash at the start of each line:

- Did you film it?
- Yesterday, at the beach.

Captions out of sync

SRT and VTT times match the video as it was posted. If you trimmed the start in your editor, every caption is off by the same amount. Some captions also stay on screen through a pause, because the cue ends where the model decided the phrase ended.

Fix: select all the captions in your editor and move them together. Shorten cues that hang over silence. How to make an SRT file has the editor steps.

Punctuation and numbers

Punctuation is added automatically. Sometimes it merges two sentences or drops a question mark. Numbers can come out as digits or words, like “25” or “twenty-five”, depending on how they were said.

Fix: read the text once before you publish. A 3-minute transcript takes a couple of minutes to proofread.

Getting better results on your own videos

• Keep the mic close. A clip-on mic, or a phone at arm’s length, beats a phone across the room.
• Keep background music low under your voice.
• Avoid wind and echo. A small room with soft furniture sounds better than a kitchen.
• If the detected language is wrong, fix it right away with Another language.

FAQ

Does trying again count toward my limit?

Not in the same language: the saved transcript comes back without using your quota. Changing the language with Another language makes a new transcript. Your first language change for each video is free that day; further changes of the same video count as one each. Attempts that end in an error don’t count.

Why did I get the same transcript again?

Transcripts of public videos are kept for up to 7 days by video ID. Transcribing the same video in the same language returns the saved text. If the language was the problem, use Another language. Otherwise, edit the text by hand.

Will an HD version of the video help?

Only if the audio is better. Video quality doesn’t change the sound, and a repost usually has the same audio as the original, or worse.

The audio is clear and the transcript is still wrong. What now?

Email [email protected] with the link and a short note on what went wrong. We use these reports to check the tool.