What is ASR hallucination? It is text that a speech recognition model produces although no one spoke it. The model fills a gap with a plausible phrase, most often over silence or music. The words read fine, which is why the error is easy to miss.
Transcribe a recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram
Files up to 2 GB · export to TXT, MD, DOCX, PDF, SRT and VTT · speaker labels and timestamps · about 30 languages
Key facts
- The invented lines are usually fluent and short, so they pass a quick read and need a comparison with the audio.
- Stretches with no speech are the typical trigger: a pause before a meeting starts, an intro jingle, or a long hold tone.
- Repeated phrases that loop several times in a row are another sign that the model lost track of the sound.
- Cutting dead air from a recording before transcribing removes many of the openings for this error.
What it looks like
[00:00:00 --> 00:00:04] Welcome to the weekly review.
[00:00:04 --> 00:00:19] (music only in the recording)
[00:00:19 --> 00:00:23] Thank you for watching. <-- invented line, nobody said it
[00:00:23 --> 00:00:30] Let's start with the budget.
More
See the full list of languages, the tools and the glossary.
Frequently asked questions
Is a hallucination the same as a misheard word?
No. A misheard word replaces something that was said. A hallucination adds words where nothing was said, so there is no spoken source to compare it with.
Where do hallucinations appear most often?
They appear on silence, music, background hum and the quiet seconds at the start or end of a file, where the model has little real speech to follow.
How can I spot invented text in a transcript?
Play the audio at the timestamp of any line that seems out of place. If the track is quiet or only music there, the line was probably invented.
Can I stop it completely?
No setting removes the risk entirely. Trimming silence, using a clean recording and reading the result against the audio reduce it, but a final check stays necessary.
Does the Potok bot check for this?
The bot returns the transcript as recognised, so read it with the audio, especially around pauses and music. You can then fix lines in an exported TXT or DOCX file.
Sources
Each source was opened by hand on the date shown. Pages change, so open the source if a detail matters to you.
- Speech recognition (Wikipedia), Wikipedia · checked
- Whisper speech recognition system (Wikipedia), Wikipedia · checked
Try it on your own recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram