You ran a recording through an AI transcriber and the transcript is mostly clean, except every name is wrong. Your CEO became "Stephanie," the city of Worcester became "Wooster," the doctor named Sirisha shows up as "Sarisha" or just "[Unintelligible]." The rest of the sentence makes sense. The names don't.
This is the most predictable failure mode in AI transcription, and it's fixable in two minutes once you know what's going on.
Why does AI transcription get proper names wrong?
Speech-to-text models don't know names. They learned to predict the most likely word given the surrounding audio and context, weighted by how often each word appears in their training text. "Stephen" appears far more often in everyday text than "Sirisha." When the model hears something between the two, it bets on common.
That bet is right for ordinary words and wrong for proper nouns by design. Add an accent, a soft consonant, or a noisy microphone, and the model leans even harder on what it has seen before — which means the rarer the name, the worse the guess.
There are three layers of fix, in order of leverage:
- Tell the model the names before it starts.
- Clean up the audio so the model has a better signal to work with.
- Post-process: do a find-and-replace pass for the names you know will appear.
The first one alone solves it most of the time.
The fastest fix: give the model the names up front
Most modern transcription engines support biasing — a way to whisper the proper nouns into the model's ear before it transcribes. The mechanism varies, but the idea is identical: you hand over a small vocabulary list, the model gives those words a higher prior probability, and the guesses snap into place.
Speakers, companies, products, places, jargon. Sloppy here means sloppy output.
In OpenAI's Whisper, this is the initial_prompt parameter: a short string of names and context that the model conditions on. AssemblyAI calls it word_boost. Deepgram exposes it as keywords. Most paid APIs have some equivalent.
The rare names should settle into their correct spelling. If they didn't, your list was missing the form the speaker actually used (e.g. "Steph" vs "Stephen").
Write names the way they're spoken, not always the way they're spelled. If your CFO goes by "AJ" in meetings but signs emails "Andrew James," put "AJ" in the bias list.
Cap the bias list. Long lists hurt more than they help: the model spreads probability across all of them. Keep it to the names that actually appear in this recording, not every name in your company directory.
What are the most common causes of name errors?
Even with biasing, names can still come out wrong. Here are the failure modes ordered most to least common.
The model never saw the name (rare or non-English origin)
Names like Aoife, Siobhán, Olufemi, or Xochitl rarely appear in English training text. The model has nothing to anchor to and falls back to whatever phoneme cluster sounded closest. A bias-list entry with the correct spelling fixes most cases. For very rare names, also include a phonetic hint: "The host's name is Aoife (pronounced EE-fa)." That gives the model a sound-to-spelling map for the recording.
The audio is muffled exactly where the name lands
People say their names softer than they say verbs. Combine that with a far-field microphone, a sibilant S, or a child in the background, and the model is guessing from rubble. Names are usually where bad audio shows up first; the rest of the audio quality fundamentals covers what to fix at the recording stage.
Two speakers introducing themselves over each other
The classic Zoom-call moment: "Hi, I'm—" "Yeah, hi, I'm—". The model collapses two names into one or attributes a name to the wrong voice. Speaker separation is a different problem from name accuracy. If this is biting you, look at speaker diarization: diarized output makes it easy to find and fix cross-attributed names.
Homophones the model resolves toward the common word
"Mr. Reid" becomes "Mr. read." "Dr. Wright" becomes "Dr. right." The model prefers the dictionary word over the proper noun because both are valid English. Fix: a bias-list entry with the title attached, e.g. "Mr. Reid", "Dr. Wright".
The speaker mispronounced or shortened the name
The model can only transcribe what it heard. If "Catherine" was actually said as "Cath," expect "Cath" in the transcript. That isn't an error: it's a faithful transcription you may want to normalize in post.
How do I fix uncommon or foreign names?
For non-English names, three moves stack:
- Spell it the way the person spells it. If they write "Sıla" with a dotless i, put that exact form in the bias list. Some models drop diacritics on output — that's a normalization pass, not a transcription pass.
- Set the language hint correctly. Mixed-language audio confuses the model and amplifies name errors. For Whisper this is the
languageparameter; AssemblyAI and Deepgram expose equivalents. - Add a pronunciation gloss in the prompt. "The interviewee is Xochitl (pronounced SO-chee)" gives the model a phonetic-to-spelling map for that recording.
If you're transcribing across multiple languages in one file, name accuracy is one of several problems. See transcribing multilingual content for the broader workflow.
What if the audio itself is the problem?
Sometimes the bias list is right and the transcript still mangles names. That's usually an audio problem, not a model problem. Things to check:
- Recording level. If introductions happened before levels were set, the names landed in the quietest part of the file.
- Microphone distance. Lavalier mics catch names cleanly. Conference-room ceiling mics don't.
- Compression artifacts. Re-encoded WhatsApp audio or low-bandwidth Zoom strips the high frequencies that distinguish similar consonants.
If you can get the speakers to redo a clean intro segment, do that and stitch it. Cheaper than fighting the transcript.
Can AI label which name belongs to which speaker?
Speaker labeling and name accuracy are related but separate. Diarization tells you that speaker A and speaker B are different people; it doesn't know either name. To turn "Speaker 2 said…" into "Maria said…", map the labels by hand once at the start of the transcript, then a global replace handles the rest.
This is also why one early bad attribution snowballs. If the model labels the wrong voice as Maria in the first minute, every "Maria said…" downstream is wrong. Scan the first two minutes carefully before trusting the rest.
Paste any public link or upload a file and get a clean transcript in minutes. First 3 clips every month are on us — no card required.
When should I give up and have a human transcribe it?
Bias lists, clean audio, and a post-processing pass fix the vast majority of name errors. The rest are recordings where the name itself is genuinely ambiguous in audio: three siblings named Steve, Stephen, and Stefan in the same family interview, a speaker who switches between formal and casual versions of their own name, or a name no published transcript has ever contained.
For those, treat AI as a first draft and have a human pass through. If you just need clean names in a clean recording, run your file through a transcriber that supports a vocabulary list and you're done. If the recording is messy and the names are the point of the recording — court testimony, medical histories, a family oral history — budget for a human review pass after the AI draft.
For what kinds of accuracy are realistic before you start, see transcription accuracy: what to expect.
