Whisp

Whisp is a lightweight desktop speech-to-text tool designed for simplicity and efficiency. Modern speech-to-text models like OpenAI's Whisper offer impressive accuracy and represent a significant improvement in usability over older technologies. Whisp aims to provide a minimal, unobtrusive interface to these models, allowing seamless speech-to-text functionality across all your desktop applications.

Goals

These are aspirational, as no code has been written yet at the time of this README.

1. Robust: Broken tools are not useful. Whisp is stable and reliable. Errors are handled gracefully, and retries are automatic when possible.

2. Minimal: Resource intensive tools with many features and poor execution are not useful. Whisp will be lightweight and unobtrusive. It will do this one thing, and do it well.

3. Secure: Handling API keys, voice data, and registering a global hotkey to enable recordings are not only security concerns but also require a lot of trust in a tool from the internet. Whisp handles your data with best practices, but in terms of trust, the best you can do is review the source and build it yourself.

Status

It is working, but bare bones. Using day to day to gather feedback and improve in free time. Currently it only supports OpenAI's api.

Basic configuration in whisp.toml:

Hotkeys are really bad right now, need to make custom hotkey parser.

hotkey = "shift+super+Semicolon"
openai_key = "not-a-real-key"
language = "en"
model = "whisper-1"
restore_clipboard = true
auto_paste = false

TODO

• Add indication when mic data starts being recieved. Currently difficult to know when exactly you can start speaking. Would like to have this indication be a visual in the bar.
• Notifications on errors.
• Basic settings page
• Document settings
• Add post processing
• Add audio pre-filter to remove dead air (saves money, smaller uploads etc.)
• Very difficult to test, need to figure out how to do this.
• Set up basic CI/CD, releases, etc.
• Support other models/APIs