Skip to content

Feature Request: Custom Vocabulary Support for Parakeet and Parakeet Ultra #127

Open

Description

@nukeador

Description

Currently, Decktation supports custom recognition keywords for Whisper through its initial prompt. It would be useful to extend this functionality to Parakeet and Parakeet Ultra.

This is particularly important for gaming, where character names, locations, abilities, items, and other game-specific terminology can be difficult to recognize, especially when mixing languages.

Proposed implementation

Reuse the existing recognition keywords and game profile configuration for Parakeet, without introducing new dictionary formats or additional user configuration.

While Parakeet does not support Whisper-style prompts, it can support custom vocabulary through native word and phrase boosting during decoding.

There is already an upstream implementation for this in "transcribe.cpp":

handy-computer/transcribe.cpp#221

This PR adds phrase boosting to Parakeet using the existing "run_params.vocabulary" API, making it possible to pass a vocabulary directly to the inference engine.

Expected behavior

  • Reuse the same recognition keywords currently used by Whisper.
  • Support both Parakeet v3 and Parakeet Ultra.
  • Support multilingual vocabulary, including foreign and fictional names.
  • Improve recognition of custom words and phrases without restricting normal speech recognition.
  • Keep processing fully offline with minimal performance overhead.
  • Avoid additional models, heavy dependencies, or changes to the existing dictionary format.

Alternative

Handy implements custom vocabulary through fuzzy and phonetic matching after transcription. While this could be a simpler fallback, native phrase boosting would be preferable because it influences recognition directly and does not rely on language-specific phonetic matching.

Handy's implementation:
https://github.com/cjpais/Handy/blob/main/src-tauri/src/audio_toolkit/text.rs

Recommendation

Evaluate the upstream phrase boosting implementation and integrate it once available in "transcribe.cpp".

Ideally, custom vocabulary should work transparently across all supported speech recognition models, using the existing Decktation configuration without requiring users to manage separate dictionaries.

References

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions