Using VoiceFlow

Custom dictionary

Teach the transcriber the proper nouns and jargon you use, so names and technical terms come out spelled correctly.

Storage

One JSON file per word

Sent to Whisper

Up to 100 terms

Scope

Transcription prompt only

Why a dictionary helps

Whisper transcribes phonetically and will guess at unfamiliar names. Dictionary terms are folded into the transcription prompt as preferred spellings, nudging the model toward the right form for proper nouns and domain vocabulary — product names, people, libraries, acronyms.

Terms are a hint, not a hard filter. The prompt asks the model to prefer these spellings when they match the speech; it won't force a term onto audio that doesn't contain it.

Adding and removing terms

  • Add a word or short phrase from the Dictionary screen. Each entry is saved as its own JSON file with an id and timestamp.
  • Duplicate entries are detected and skipped.
  • Delete a term at any time; it stops influencing new recordings.
  • The first 100 non-empty terms are included in the transcription prompt for each recording.

The hallucination guard

A known failure mode of prompt-based biasing is that the model echoes your dictionary terms even when you didn't say them. VoiceFlow watches for this: if a short transcript is made up almost entirely of dictionary tokens, it is flagged as a likely hallucination in the logs. The two-pass transcription strategy also penalizes dictionary-heavy output when scoring candidates.

Keep the dictionary focused on terms you actually use often. A smaller, high-signal list biases better than a giant one.