A bad transcript usually has one or two causes, and the mistakes on the page tell you which. Seven common causes of audio transcription errors, what each one looks like, why it happens, and the fix in the recording and in the transcript.
Audio transcription errors are rarely random. A transcript that turns "Kubernetes" into "Cooper Netties", puts two people's words under one name, or writes "fifty" where the speaker said "fifteen" is showing you what went wrong. Each cause leaves its own kind of mistake, and each has its own fix. Some fixes belong in the next recording, some in the transcript you already have.
This guide starts from the mistake. Find the one that matches your transcript in the list below, then go to its section: what the error looks like, why it happens in plain words, and what to do about it.
Most transcription problems fall into seven patterns. Match the symptom to the cause:
Two causes often stack. A remote guest on a headset in a noisy café, speaking with an accent, will produce three kinds of error at once. Fix the cause that affects the most lines first: usually the audio, then the language, then the words.
What it looks like. In a recording made in a café, "Let's move the launch to March" comes back as "Let's move the lunch to March". A word under a door slam disappears. Over a long silence with an air conditioner humming, a line appears that nobody said, such as "Thank you for watching."
Why it happens. Speech recognition guesses the most likely words for the sound it receives. Noise covers the quiet parts of speech first: the ends of words and the short words between them. With less of the real signal, the guess leans more on what usually comes next, and "lunch" is a more common word than "launch". Over steady noise with no speech at all, some systems still produce a stock phrase, because they were trained to write something whenever they hear sound.
Fix it in the recording. Move the microphone closer to the speaker rather than turning it up. Doubling the distance cuts the speaker's direct sound by about 6 dB while the room noise stays where it was, so closeness matters more than any setting. Switch off fans and air conditioning for the recording if you can, and record in a room with soft surfaces rather than glass and tile.
Fix it in the transcript. In VocalStack, name the spoken language on noisy recordings instead of leaving it on automatic detection. The Dashboard's own note says so: naming the language improves accuracy on noisy recordings and heavy accents. For a live transcription, the Accuracy latency mode accepts a little more delay in exchange for fewer mistakes on hard audio. Lines the engine is least sure of (below the transcription's confidence threshold, 0.60 by default) are shown to listeners as "Audio unclear" instead of a guess, and are never read aloud. A line that is pure invention can be hidden by hand: hidden lines are left out of the transcript, exports, subtitles and the player.
What it looks like. "The data shows a thirty percent rise" becomes "The data shows a dirty percent rise". The mistakes follow one speaker through the recording, while everyone else's lines are clean.
Why it happens. A recognition system learns how words sound from the speech it was trained on. When a speaker pronounces a sound differently from most of that speech, such as "th" said as "d", the system hears exactly what was said and picks the real word that matches that sound best. Automatic language detection can make it worse: a strong accent can shift the cues towards another language altogether (see how automatic language detection works).
Fix it in the recording. For a live talk, add a Reference Script before you start (see the jargon section below), so the words the speaker is likely to say are already known. A good microphone helps every accent, because more of each sound reaches the engine.
Fix it in the transcript. Name the spoken language for that recording. Then use the Fix tab: Mark uncertain words colours and underlines the words the engine was unsure of, and you can step through the uncertain segments one at a time instead of rereading everything. Where the engine considered other words, it offers them as suggested fixes, so correcting "dirty" to "thirty" is usually one click. Play just this word lets you check by ear before you change it.
What it looks like. One speaker asks "So we agree on Tuesday?" while another cuts in with "No, Thursday works". The transcript shows a single line under Speaker 1: "So we agree on Tuesday no Thursday works." Or the opposite: one person who got animated halfway through the meeting appears as Speaker 1 and Speaker 3.
Why it happens. Two separate jobs run underneath. Speech recognition writes the words; speaker diarization decides who spoke when, from how each voice sounds. Most systems give each stretch of audio one speaker, so when two people talk at once, one of them loses. A change in someone's voice (excited, tired, further from the microphone) can make them look like a new person. Speaker diarization vs speech recognition explains both jobs in detail.
Fix it in the recording. Give each person their own microphone. That one change does more for multiple speakers than any setting. At a panel or a meeting with audience questions, ask the moderator to repeat each question into their own microphone before it is answered.
Fix it in the transcript. Choose the speaker labels setting: When Needed (the default) labels lines once a second voice is heard, On labels every line, Off labels none. Name each speaker during or after the recording. The name is a layer over the words, so it never changes the text, and it applies to that person in every recording of the transcription. Save a speaker you record often, and later transcriptions recognise them by name. A line that runs two sentences together can be split in two: tap the gap where the new line starts.
What it looks like. "We deployed it on Kubernetes" becomes "We deployed it on Cooper Netties". "Ms Nguyen" becomes "Miss Win". A drug name turns into three short English words. The rest of the sentence is fine.
Why it happens. A system can only write words it knows, and a name often appears once, with nothing around it to predict it from. When it has never seen the word, it picks the ordinary words that sound closest. This is the hardest case for accurate transcription of technical terminology: the sentence gives no clue, so no amount of context fixes it.
Fix it in the recording. Give the system the words before the speaker says them. In VocalStack that is a Reference Script: add the script or slides of a live talk, or a list of names and terms, before you start, and names and terms from it win over what was misheard. During the talk the script follows the transcript like a teleprompter. Every word the script corrected is marked SCRIPT among the fixes and can be undone if the speaker really did say something else.
Fix it in the transcript. Correct the name once and check the rest with search. Corrections reach every listener while the transcription is live, in every language they are reading, and every correction is listed and can be undone. Exports (SRT, PDF, Excel, JSON) include the corrections, so subtitles and documents carry the fixed spelling.
What it looks like. A talk opens with a greeting in Spanish and continues in English, and the English comes out as Spanish-looking nonsense. Or a few lines in the middle of an English transcript appear in a related language. Some systems, set to the wrong language, write a rough translation instead of what was said.
Why it happens. When no language is named, the system has to detect it from the sound, often from the first few seconds. A greeting, a quotation or a strong accent can point it the wrong way. Once it has chosen the wrong language, every word in that stretch is wrong together.
Fix it in the recording. If the recording is in one language, name it. Leave detection on when two or more languages are really spoken, so each part is transcribed in its own language; naming one would force the other into it. In a live transcription the language can be changed while it runs, and the change applies at once.
Fix it in the transcript. A line heard in the wrong language cannot be fixed word by word, because nothing in it is right. VocalStack can transcribe the line again: pick the language it was really spoken in, and the line's audio is read again in that language. The result replaces the line as an ordinary correction, with the same undo as any other.
What it looks like. "The fees went up" becomes "The thieves went up". "Free" and "three" swap. The errors spread across the whole recording, and they are worst for the remote speaker on a call.
Why it happens. Many of the sounds that tell "f", "s" and "th" apart sit high in the frequency range. A phone line carries nothing above about 3.4 kHz, and a Bluetooth headset often switches its microphone to a call mode with a much narrower range. Low-bitrate compression in video calls removes more. The system receives audio in which those sounds are already gone, and it guesses between words that now sound identical.
Fix it in the recording. Use a wired microphone, or a headset whose microphone stays in its full-quality mode. Record each remote speaker on their side if you can, rather than recording the call. For an event, take the feed from the audio mixer instead of a laptop or phone microphone in the room: VocalStack's live setup takes one cable from the mixer, and a phone's built-in microphone picks up more of the room.
Fix it in the transcript. Audio that was never captured cannot be recovered, by a machine or a person. Use the uncertain-word marks to find the damaged stretches, play the line to check, and correct what you can hear. Where nobody can tell what was said, hiding the line is more honest than leaving a guess.
What it looks like. "Fifteen" becomes "fifty". "One twenty" becomes "120" when the speaker meant one dollar twenty. "The third of the fourth" comes out as a date in the wrong order. Each mistake is one word, and every other word in the sentence is right.
Why it happens. "Fifteen" and "fifty" differ mainly in which syllable is stressed, and that difference is the first thing lost to noise or a fast speaker. Spoken numbers also leave out what written ones need: units, decimal points, the order of day and month. The system has to guess the format, and its guess is the most common one, not yours.
Fix it in the recording. Ask speakers to say figures in full, with the unit: "one dollar twenty", "the fifteenth of March". For a prepared talk, put the figures in the script.
Fix it in the transcript. Check every figure that matters by ear before the transcript is quoted, sent or published. A wrong number reads like a correct one, so nothing on the page warns you. What word error rate measures, and what it misses explains why a single wrong figure costs more than its share of the error rate.
When a transcript has many problems, fix them in this order, so that no fix is undone by the next one:
Then make the next recording easier: a microphone per speaker, the language named, and a script with the names in it. Most audio transcription errors are decided before anyone presses record.
For a hearing, a medical consultation or anything where one wrong word has consequences, have a person check the transcript before it is used, whichever system made it. To see which of these errors your own audio produces, try a recording in the Dashboard.