Send custom vocabulary as biasing prompt to custom transcription endpoints

[tejadhith]

Problem

The custom dictionary (Vocabulary) is passed to the built-in cloud providers, but the custom OpenAI-compatible path never receives it. Users running models like gpt-4o-transcribe on custom or Azure OpenAI endpoints get no transcription-time term biasing — every technical term, product name, or proper noun in their dictionary is ignored until post-hoc word replacement, which can't fix words the model never produced.

Change

• New shared helper VocabularyBiasingPrompt builds a framed instruction from the vocabulary — "The audio may reference the following names, products, and technical terms — transcribe any occurrences using these exact spellings: …" — rather than a bare comma list (a bare list gives instruction-following transcription models no guidance on what to do with it). Terms are trimmed, case-insensitively deduped, and capped at 200 to stay inside prompt budgets.
• OpenAICompatibleTranscriptionService emits the result as the multipart prompt field — only when there is something to send.
• CloudTranscriptionService passes the dictionary terms into the custom-model path (it already fetched them for built-in providers).

The helper is deliberately shared so other providers can adopt the same biasing format.

Relationship to the prompt-scoping change

This intentionally does not reintroduce the user's free-form TranscriptionPrompt on custom endpoints: cde93d3 ("Scope transcription prompts to Whisper models") nils context.prompt at the source for non-Whisper models, and that stays exactly as is. What's sent here is a different, narrower thing — a short app-constructed instruction derived only from the user's Vocabulary entries; with an empty dictionary the field is not sent and the request is byte-identical to today's.

prompt is part of the OpenAI transcription API spec, so OpenAI-compatible endpoints are expected to accept (or ignore) the field; if you'd rather make this opt-in per custom model, happy to add a toggle.

Testing

Verified against an Azure OpenAI gpt-4o-transcribe deployment: a product name the model consistently mis-heard is transcribed with the exact expected spelling once added to the Vocabulary, and the empty-dictionary case was confirmed to send no prompt field.

---

Summary by cubic

Pass custom vocabulary to OpenAI-compatible transcription endpoints as a biasing prompt so custom models like gpt-4o-transcribe use exact term spellings during transcription. No prompt is sent when the dictionary is empty.

• New Features
  • Added VocabularyBiasingPrompt to build a short instruction from trimmed, deduped terms (max 200).
  • OpenAICompatibleTranscriptionService sends it as the multipart prompt when non-empty; otherwise unchanged.
  • CloudTranscriptionService now forwards vocabulary to custom models; this does not re-enable free-form user prompts for non-Whisper models.

^{Written for commit 92c5d98. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}