Dictionary Replacement

Dictionary replacement is a rule-based cleanup step that converts known words or phrases into preferred output after the speech recognition pass.

Why it exists

Recognition models work probabilistically. They try to infer the most likely words from the audio signal. That is useful, but it also means product names, brand names, and internal jargon can drift. The model may hear the right sound and still output the wrong spelling.

Dictionary replacement handles that gap with deterministic rules. If a known phrase appears, it is rewritten into the exact target form you want.

What it is good at

• Brand names: keep product and company names consistent.
• Team vocabulary: standardize internal terms that are easy to misspell.
• Mixed-language writing: protect English terms inside Korean dictation and vice versa.
• Cleanup discipline: reduce the same repetitive corrections every day.

Why deterministic matters

Deterministic cleanup is valuable because it is predictable. Unlike a model that may vary from one pass to the next, the same rule yields the same output each time. That makes it easier to trust in high-frequency workflows where users want stable text, not creative variation.

In Mallo, dictionary replacement is one of the clearest ways to close the gap between "pretty good recognition" and "ready to send text."

Common mistakes

• Overloading the dictionary: if every possible phrase becomes a rule, maintenance gets messy fast.
• Using rules to hide model issues: dictionary replacement is best for known vocabulary, not for fixing every broader recognition problem.
• Ignoring context collisions: aggressive replacements can create false positives when a phrase appears in an unrelated sentence.