You have 20 finished videos and the auto captions in your editing software keep turning gfx1201 into “GFX 1201” and vLLM into “VLLM”. The fix is not a better font of patience — it is giving the speech model a custom word list before it transcribes. This page explains how that works and how this tool does it.
Why auto captions fail on jargon
Speech recognition models pick the word sequence that is most probable for the audio they hear. Everyday words are high probability, so they survive. Rare tokens — acronyms, version numbers, model names, product names, internal tooling — are low probability, so the model quietly swaps them for the nearest common word. Nothing about your audio is wrong; the model simply has never heard your vocabulary.
That is exactly why “fix captions more than I would writing them” is such a common complaint in technical content: every mangled term is a manual edit, and every video has dozens.
How a word list fixes it
Whisper-class models expose a context hook — an initial_prompt (sometimes called hotwords or vocabulary bias) — where you can put the terms you expect to hear. The model raises the probability of those tokens while it decodes, so when it hears “V-elle-ell-eme” it writes vLLM the way you spelled it, not the way a general model guesses.
The list is a bias, not a find-and-replace: it does not inject your words into the transcript, it just makes the model far more likely to land on them when the audio supports it.
Using it here
- Go to the free tool and use the local-file upload — MP3, WAV, M4A, MP4, MOV, up to 500 MB.
- In the Custom word list box under the upload, paste the terms that keep getting mangled — one per line, or comma-separated. Your highest-frequency terms first.
- Upload. The transcription runs with your list applied, and the SRT / VTT / TXT output already has the right spellings.
- Create a free account and press Save my word list — every future upload uses it automatically, and the per-file box is just for one-off extras.
What works best in a word list
- Exact spellings — the casing you want in the output. “PyTorch”, not “pytorch”, if that is how you write it.
- Your top 10–30 terms, not every word you ever use. A focused list beats an exhaustive one.
- Things that are genuinely rare — model names, CLI flags, product names, acronyms your audience knows and the model has not heard.
- Phrases work too: “tensor parallelism” as one entry keeps the phrase intact.
Honest limits
If a word is muffled, shouted over music, or spoken in a heavy accent, no word list fully rescues it — the audio itself is ambiguous. What the list does is remove the systematic class of errors (the ones that happen on every occurrence of a rare word), which is where almost all of the editing time goes. Budget a skim, not a rewrite.
Related
The word list works on the local-file upload. For a YouTube video with no captions at all, the transcriber handles it — see transcript for a video without subtitles. And when the transcript is the start of an AI workflow, see transcripts for LLMs (REST + MCP).
Frequently asked questions
What is a custom word list in transcription?
A list of the exact terms in your content — product names, API names, acronyms, niche jargon — that you give the speech model before it transcribes. Whisper-class models expose this as an initial prompt (or hotwords): it biases the model toward your exact spelling, so “vLLM” stays vLLM instead of becoming “VLLM” or three separate words.
Why do auto captions get technical terms wrong?
Speech models are trained on general conversation. Rare tokens — version numbers, model names, internal tooling — have low prior probability, so the model picks the nearest common words. A word list raises the prior for your terms without touching the rest of the transcript.
Does a word list guarantee correct spelling?
No — it is a strong bias, not a hard rule. If a term is genuinely hard to hear in the audio (muffled, fast, overlapping speech) it can still come out wrong. In practice it removes the majority of jargon fixes, which is where most of the caption-editing time goes.
How long can my word list be?
Around 50 terms (roughly 600 characters) is the cap on uploads; the model's own context window is the real limit, and a focused list of your highest-frequency terms outperforms a giant one. One term per line, or comma-separated.
Do I have to re-enter my word list for every file?
No. Create an account and the “Save my word list” button stores it with your profile — every upload after that uses it automatically. You can still add extra terms for a single file without changing the saved list.
What output formats do I get?
TXT with per-line timestamps, SRT and VTT subtitle files with frame-accurate timing, and JSON with per-segment start/duration for pipelines. The same formats work for the word-list transcriptions as for everything else.