Skip to content

Roadmap & Contribution Guide

Jump to bottom
Zach Nagengast edited this page Jan 30, 2024 · 2 revisions

WhisperKit is looking forward to your contributions! We have purposefully released at pre-feature complete stage to build this project with the open-source community and stress test how extensible WhisperKit really is.

Concretely, we put TODO stubs for Whisper functionality that exists in the OpenAI reference Python implementation but not yet in WhisperKit. Here is the current list to reach parity:

  • Suppress Tokens
  • No speech thresholding
  • Word-level timestamps
  • Beam Search

On the other hand, we are actively working towards adding the following features that are more involved to contribute because they require improvements to the models themselves to be efficiently supported:

  • Multi-token TextDecoder forward pass for <|startofprev|>-based prompting (Current TextDecoder can only decode auto regressively)
  • Speculative Decoding
Clone this wiki locally