Skip to content

README.md

Latest commit

 

History

History
55 lines (37 loc) · 3.51 KB

README.md

File metadata and controls

55 lines (37 loc) · 3.51 KB
Raw

<- Back to Main README

Model Management Feature Plan

This document outlines the high-level tasks for implementing a model management feature within the WhisperPrototype application. This feature will allow users to manage the Whisper model files used by the application.

See also: Roadmap

Current Model Handling & Motivations for Change

  • Models are currently stored in a ./Models subdirectory relative to the application's execution path (AppContext.BaseDirectory).
    • Issue ⚠️ If models are included in the project and set to "Copy to Output Directory", this can lead to large model files (often 1-3GB each) being duplicated in the build output for each configuration (e.g., Debug, Release). This consumes significant disk space and is inefficient.
  • Model selection occurs at application startup via a CLI prompt in Workspace.cs.
  • The download-ggml-model.sh script provides a reference for model names, sources, and download logic.
  • Goal: Transition to a system where models are stored in a single, canonical location in userspace (user-configurable), preventing duplication in build outputs and ensuring all parts of the application reference this single store.

Task Breakdown

  1. Configure Model Storage Location

    • Allow users to specify a custom directory for storing models.
    • Configuration persisted in a user directory dotfile i.e. ~/.config/WhisperSpeechToTextDotnet (and the Windows 11 equivalent, if run via PowerShell).
    • Update Workspace.cs to use this configured path. If not set, default to the current ./Models path.
    • Display the currently configured model storage location to the user within the "Manage Models" section.

  2. Repeat (1) for Inputs

    • Repeat the above steps for an additional configuration value, which only takes care of configuring the Inputs directory, this is where the audio recordings live.

  3. List Available and Downloaded Models

    • List Local Models: Scan the configured models directory and display a list of already downloaded/present .bin files.
    • List Remote Models:
      • Fetch or maintain a list of available models (similar to the list in download-ggml-model.sh). This could be from a hardcoded list, a simple config file, or (more advanced) by querying a source like Hugging Face if feasible and desired.
      • Indicate which remote models are already downloaded.

  4. Download New Models

    • Allow users to select a model from the list of available (not yet downloaded) remote models.
    • Implement download logic within the C# application.
      • This could involve making HTTP requests to the model URLs (e.g., from Hugging Face, as seen in the shell script).
      • Show download progress if possible.
      • Save the downloaded model to the configured models directory with the correct naming convention (e.g., ggml-<model_name>.bin).
      • Perform a checksum verification after download if feasible (SHA hashes are available in the README.md model table).

  5. Delete Existing Models

    • Allow users to select one or more downloaded models from the local list.
    • Confirm deletion with the user.
    • Delete the selected model file(s) from the models directory.

  6. User Interface (within "Manage Models" section)

    • Use Spectre.Console components (SelectionPrompt, MultiSelectionPrompt, etc.) for interactions.
    • Provide clear feedback on actions (downloading, deleting, listing, errors).