Guide

Why transcription gets names wrong, and what to do about it

Every speech model is worse at names than at ordinary words, and the reason explains how to work around it.

5 min read · Updated 10 September 2026

Names are the hardest words in any language

A speech model predicts the next word from context, and context is exactly what a proper noun lacks. “I spoke to…” could be followed by any name on earth, so the model falls back on what it heard acoustically — and acoustically, Rahul and Raoul are nearly identical.

This gets worse in a room like ours. A transcript that has to hold Arabic, South Asian, Filipino and European names is being asked to distinguish sounds that several of those languages do not treat as different sounds at all.

Say the name early, and say it with a role

“Ahmed, you were going to speak to the contractor” gives the model two things: an acoustic sample of the name and a context that makes it a person. A name said once in isolation, in the first ten seconds, before anyone has settled, gives it neither.

The practical version of this is the round of introductions everyone skips on a call with people who already know each other. Doing it anyway is thirty seconds that make the whole transcript readable.

Spell the ones that matter

If a name is going into a contract or a medical record, spell it out loud once. It feels stilted and it is the difference between a record you can rely on and one you have to check.

The same applies to company names, project codenames and anything invented. A model has never heard your internal project name and will render it as the nearest real word every time.

Fix it once, not every time

Most transcription tools, ours included, let you add terms the model should prefer — the names of your team, your clients, your products. Ten minutes of setup removes the most common error permanently.

This is the single highest-return thing you can do with any transcription tool, and almost nobody does it.

Know what a wrong name costs

In a personal note, nothing. In a clinical record or an attendance note, a great deal — which is why a transcript is a draft you review, not a document you file. Any tool that encourages you to skip the review is selling you a risk, ours included if we ever did.

Keep reading

How to record a conversation without making it strange

The awkwardness of recording is not a technical problem, and no device solves it for you. This is what actually works.

5 min read →

Running a meeting that switches between Arabic and English

Code-switching is not a problem to be fixed. It is how business here is conducted — and most note-taking tools were not built for it.

6 min read →

What to actually do with a meeting transcript

A transcript is not notes. Nobody reads an hour of raw text, and treating it as an archive is how it becomes worthless.

4 min read →

How to choose an AI note taker without regretting it

Six questions that separate the tools that survive a month of real use from the ones that end up in a drawer.

6 min read →

Five things that ruin a meeting recording

None of them are the microphone. All of them are fixable in the ten seconds before a meeting starts.

4 min read →

Writing clinical notes without typing during the consultation

Typing changes what patients tell you. Here is how clinicians work around that, with and without a recording device.

6 min read →

Handing over a client without losing the context

The CRM has the facts. What gets lost is everything that explains them — and that is what the next person needs.

5 min read →

How long should you keep a recording?

Most people keep everything forever by default. That is a decision, even when nobody made it deliberately.

5 min read →

Running a one-to-one people do not dread

The format is not the problem. The problem is that nothing said in the last one visibly happened.

5 min read →
Chat with a human · replies in ~1h