Skip to content

[codex] Add Parakeet word boosting#142

Open
HashiamKadhim wants to merge 1 commit into
Starmel:masterfrom
HashiamKadhim:parakeet-word-boosting
Open

[codex] Add Parakeet word boosting#142
HashiamKadhim wants to merge 1 commit into
Starmel:masterfrom
HashiamKadhim:parakeet-word-boosting

Conversation

@HashiamKadhim

@HashiamKadhim HashiamKadhim commented Apr 25, 2026

Copy link
Copy Markdown

Summary

  • adds a Parakeet boosted-words editor in Model settings and persists the custom vocabulary
  • loads the vocabulary through FluidAudio custom vocabulary support when the Parakeet engine initializes
  • persists FluidAudio rescorer fixes so longer matching spans win over shorter overlapping spans and short clips do not crash when multi-word aliases are configured
  • removes the nested scroll view that was squishing the Parakeet model picker

Validation

  • git diff --check
  • xcodebuild test -scheme OpenSuperWhisper -configuration Debug -derivedDataPath build -destination 'platform=macOS,arch=arm64' -skipPackagePluginValidation -skipMacroValidation -UseModernBuildSystem=YES -clonedSourcePackagesDirPath SourcePackages -only-testing:OpenSuperWhisperTests/AppPreferencesParakeetVocabularyTests CODE_SIGNING_ALLOWED=NO CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO
  • ./run.sh build
  • signed and verified the debug app with codesign --verify --deep --strict
  • saved-audio probe with a generic boosted entry like customterm: spoken alias: base transcript matched the spoken alias, boosted transcript emitted customterm
  • short one-word saved-audio probe with multi-word boosted aliases configured returned normally instead of crashing

@HashiamKadhim HashiamKadhim marked this pull request as ready for review April 25, 2026 01:19
@HashiamKadhim HashiamKadhim force-pushed the parakeet-word-boosting branch from 5c6464c to 22eff87 Compare April 25, 2026 17:42
MaximCosta added a commit to my-monkeys/OpenSuperWhisper that referenced this pull request Jun 21, 2026
Sources the Parakeet (FluidAudio) custom-vocabulary boosting from the same
custom-dictionary terms used for Whisper's prompt-boost — one list, both
engines (CustomDictionary.boostTerms is the single source). Drops the separate
boosted-words editor/pref in favour of the unified Custom Dictionary.

Includes the FluidAudio vocabulary rescorer patch (prefer longer matching
spans), applied idempotently by run.sh against FluidAudio 0.11.0.

Based on PR Starmel#142 by Hashiam Kadhim.

Co-authored-by: Hashiam Kadhim <hashiam@squareup.com>
MaximCosta added a commit to my-monkeys/OpenSuperWhisper that referenced this pull request Jun 21, 2026
Integrates with original-author credit:
- universal custom dictionary (Starmel#149, AlexCherrypi): prompt-boost (Whisper) +
  whole-word post-transcription replacement, on both engines.
- Parakeet decode-time boosting sourced from the same dictionary (Starmel#142,
  Hashiam Kadhim), unified into a single editor; FluidAudio vocabulary
  rescorer patch (prefer longer matching spans), applied idempotently by run.sh.

Stays on FluidAudio 0.11.0 (the version with custom-vocabulary support; 0.15.x
removed it — upgrade is a separate future task). All tested (101 unit tests
green, 4 pre-existing env failures), build green, GUI/voice verified.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant