desk-talk

Transcription for your desktop.

A modern GUI application that records what you say when you press a button down, and types what you said when you release it.

Important

⚠️ This video contains sound and is intended to be listened to with audio on. ⚠️

transcribe.mp4

Features

✨ Push-to-Talk Transcription - Hold a key, speak, release to paste
⚡ Realtime Streaming - Types text live as you speak via gpt-realtime-whisper
✒️ Smart End Punctuation - An AI picks the correct ending mark (. ? !, language-aware)
🎯 System Tray Integration - Runs quietly in the background
⚙️ Modern GUI Settings - Beautiful, easy-to-use configuration interface
📊 WPM Statistics - Track your words per minute with rolling averages
🔒 Secure API Storage - API keys stored safely in Windows Credential Manager
🎤 Multiple Audio Devices - Choose any microphone on your system
🌐 OpenAI or Local - Use OpenAI's Whisper API or run models locally
💾 Persistent Settings - All preferences saved between sessions

Setup

Make sure ffmpeg is installed and added to your PATH

Quickstart

1. Download and run the latest release
2. Click the tray icon in your system tray to open settings
3. Configure your settings:
   • Set your push-to-talk key (e.g., Scroll Lock)
   • Choose your microphone
   • Enter your OpenAI API key OR enable local transcription
4. Click "Start Transcription"
5. Hold your PTT key, speak, and release!

Note

You can manage your OpenAI API keys at https://platform.openai.com/api-keys

Using the GUI

System Tray

• Left-click the tray icon to open the settings window
• Right-click for quick Start/Stop/Quit options

Settings Tabs

General Tab:

• Configure push-to-talk key
• Select audio input device
• Toggle capitalization and spacing options
• Choose between paste mode (default) or typing mode

Transcription Tab:

• OpenAI Mode: Enter your API key for cloud transcription
• Local Mode: Download and run Whisper models on your computer
  • Available models: tiny-en, tiny, base-en, base, small-en, small, medium-en, medium, large-v1, large-v2, large-v3
  • Larger models = better accuracy but slower processing

Statistics Tab:

• View your current words per minute
• Track rolling average WPM (last 1000 samples)
• Monitor total words transcribed
• See total recording time

WPM Display

After each transcription, the console displays:

WPM: 132.4 (27 words over 12.25s) | Avg: 118.7

• Current WPM - Speed of this transcription
• Word count - Number of words spoken
• Duration - Time from key press to release
• Rolling average - Average of your last 1000 transcriptions

Transcription modes

DeskTalk has two ways to turn your speech into text. Pick one in General → Transcription Mode. The out-of-the-box default is Realtime (with xhigh accuracy and smart end punctuation) so it "just works" with the least fuss; switch to Standard or dial settings down if you want to save on API cost.

Standard (record, then transcribe)

Records your audio to a temporary file while you hold the key, then transcribes the whole thing once you release. Highest fidelity, since the model sees the entire recording at once. Supports Parallel Racing (send multiple requests at once and use the fastest/most reliable result).

Realtime (types as you speak)

Streams your microphone to OpenAI's gpt-realtime-whisper model over a WebSocket and types the words into the focused window live, as you talk — much lower perceived latency and a visual sense of progress. Requires the OpenAI API (not available in local mode).

• Accuracy / latency is tunable with the delay setting: minimal → low → medium → high → xhigh. Higher gives the model more audio context before it commits text (better accuracy, text trails a bit further behind your voice). Default: xhigh.
• A debug log of each realtime session is written next to your config at %APPDATA%\desk-talk\desk-talk\config\realtime.log.

End punctuation

A single setting controls the mark at the end of each utterance (the old mutually-exclusive "always end with period" and "smart punctuation" toggles are now one choice):

• None – leave the ending exactly as transcribed.
• Period – add a plain . if it doesn't already end with ., ?, or !.
• Smart (default) – a cheap, fast model (gpt-4o-mini) picks the correct ending mark for what you said — ., ?, !, or the appropriate mark for your language. It works in both transcription modes and requires the OpenAI API.

Smart mode skips the LLM call entirely when the text already ends with a terminal mark — detecting that punctuation is present is a trivial local check and needs no model — so you only pay for an API call when a mark is actually missing.

Command-line flags

The GUI app reads a few flags at launch that override the saved settings for that run (handy for launchers/scripts). If a flag isn't passed, the saved setting is used.

Flag	Description
--realtime / --no-realtime	Force realtime streaming on / off
--realtime-delay <level>	minimal, low, medium, high, or xhigh
--end-punctuation <mode>	none, period, or smart
--parallel <n>	Number of parallel requests to race (Standard mode), 1–5

--period, --smart-punctuation, and --no-smart-punctuation are kept as aliases for --end-punctuation period, --end-punctuation smart, and --end-punctuation none respectively.

Example:

desk-talk.exe --realtime --realtime-delay xhigh --end-punctuation smart

Building from Source

# Install Rust and dependencies
cargo build --release

# The executable will be in target/release/desk-talk.exe