My own lil' dictation app. For details see: Making a dictation app in a weekend.

Works on my:

• M1 MacBook Air running MacOS 14.7.5 with XCode 16.1.
• M4 Mac Mini running MacOS 15.1.1 with XCode 16.2.

Run:

./build.sh

Then:

• copy the app to your /Applications/ folder
• give it accessibility permissions
• run it

Hold down Control-Shift-H, say something, release Control-Shift-H, watch text appear.

I'm open to pairing or PRs that:

• add a single preferences window that allows you to configure the global activation hotkey
• make it easier for other people to build (is it possible without XCode?)
• graphically select and download Whisper models besides the base English one
• implement live transcription with a prinicpled approah like Whisper Streaming (I assume it's good based on the video demos, but couldn't run it myself because, lol, Python dependencies)
• add a simple CLI that makes it easy to, e.g., get rough transcription of long, unchaptered YouTube videos so you can jump to parts.
• Don't add a leading space when writing into Emacs
• Optionally use larger model

I'm not interested in PRs related to:

• other operating systems or MacOS versions (I don't use them!)
• Best Practices. The only continuous integration I'm doing is using the app myself =D

Shout out to:

• ggerganov/whisper.cpp: Port of OpenAI's Whisper model in C/C++
• Espanso for figuring out how to inject text in any app

Have a great day!

Nikita contacted me with some nice stylistic changes that he vibe-coded with Claude, which I've merged in.

This should be faster than the previous version, since now Whisper.cpp uses Metal on MacOS. It feels a bit snappier on my M1 Air, so I'm happy.

func insertStringAtCursor(_ string: String) {
    let systemWideElement = AXUIElementCreateSystemWide()
    
    var focusedElement: AnyObject?
    let result = AXUIElementCopyAttributeValue(systemWideElement, kAXFocusedUIElementAttribute as CFString, &focusedElement)
    
    if result != .success {
        logger.error("Error getting focused element: \(result.rawValue)")
        return
    }

    let element = focusedElement as! AXUIElement

    var value: AnyObject?
    AXUIElementCopyAttributeValue(element, kAXValueAttribute as CFString, &value)

    if let currentValue = value as? String {
        let newValue = currentValue + string
        AXUIElementSetAttributeValue(element, kAXValueAttribute as CFString, newValue as CFTypeRef)
    } else {
        AXUIElementSetAttributeValue(element, kAXValueAttribute as CFString, string as CFTypeRef)
    }
}

trying to use the accessibility API like this doesn't work for Cursor or other apps that have custom text input thingies. I think I might need to just generate keypress events?

ahh, found something that worked in https://github.com/espanso/espanso/blob/6b380d1edd94dff6505d97039ccb59c00ae1c5f4/espanso-inject/src/mac/native.mm#L30

https://github.com/ufal/whisper_streaming

wasn't able to run faster-whisper backend because

ValueError: This CTranslate2 package was not compiled with CUDA support

uv add whisper-timestamped

uv run whisper_online_server.py --model base.en --backend whisper_timestamped

ffmpeg -hide_banner -f avfoundation -list_devices true -i "" ffmpeg -hide_banner -f avfoundation -i ":1" -ac 1 -ar 16000 -f s16le -loglevel error - | nc localhost 43007

unfortunately I kept running into this error after connecting and speaking for a few seconds. Maybe the backend is busted somehow? I'm able to capture data via ffmpeg from my mic, so that's not it.

File "/Users/dev/software/whisper_streaming/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1616, in _call_impl hook_result = hook(self, args, result) File "/Users/dev/software/whisper_streaming/.venv/lib/python3.9/site-packages/whisper_timestamped/transcribe.py", line 882, in lambda layer, ins, outs, index=j: hook_attention_weights(layer, ins, outs, index)) File "/Users/dev/software/whisper_streaming/.venv/lib/python3.9/site-packages/whisper_timestamped/transcribe.py", line 777, in hook_attention_weights if w.shape[-2] > 1: AttributeError: 'NoneType' object has no attribute 'shape'

bash ./models/download-ggml-model.sh base.en

brew install sdl make stream ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000