Replacements
Deterministic text substitutions applied to every transcript.
Replacements are deterministic text substitutions: a source string maps to its output. The source is matched literally against the transcript, so it can be anything that appears there — kanji, Latin, or punctuation, not just a phonetic reading. Unlike Custom Words, they always apply — the same input always produces the same output.
Examples
| Source | Output |
|---|---|
| github | GitHub |
| loreal | L'Oréal |
| ikea | IKEA |
How matching works
- Case-insensitive matching of the source string.
- Longest source first, so a longer phrase wins over a shorter one it contains.
- A single left-to-right scan — rules don't cascade into each other, so the result is stable and idempotent (running it again changes nothing).
When they're applied
Replacements run twice in the pipeline: once on the raw transcript before the LLM, and once after the LLM. That means your exact output survives even when LLM cleanup is on, and they also work for Profiles that don't use an LLM at all. The LLM is additionally told to preserve already-substituted terms verbatim.
Use Replacements when you need a term to come out exactly right every time — brand names, casing, punctuation. Use Custom Words to help the model hear a term in the first place.