I built an offline transcription app in an hour

I built an offline transcription app in about an hour because I wanted dictation that worked my way. The first working version did the important job: hold a key, speak, let go and put the words wherever my cursor was. I called it Speak Kune Do. Yes, there are nunchucks.

The app now runs a Whisper speech model on my laptop. Once the model and dependencies are installed, I can dictate without an internet connection, an account or a monthly subscription. That matters when I'm travelling, when the connection is poor, or when I simply want this one task to stay on my own machine.

Why build an offline transcription app?

I like good software. I also know the feeling of paying every month for a polished product while wishing one small part of it behaved differently. With dictation, I wanted a hold-to-talk key that felt natural, text placed directly into the app I was using, and control over how names and jargon were spelled. This is a different job from AI note-taking for meetings: I am dictating my own words into the document in front of me. I did not want to change my workflow to suit a vendor's idea of a typical user.

The surprise was how quickly a useful first version came together. The underlying speech model and development tools already existed. My job was to join them around the way I actually work. That shift changes the economics of a narrow tool. The question is no longer only, "Which subscription is closest?" It is also, "How hard would it be to make the exact thing I need?"

A subscription has to earn its place when the person using it can shape the alternative.

The details I would never find on a feature list

Speak Kune Do's default is simple: hold Right Alt, talk, release. It shows a rough live transcript while I speak, then makes a more careful pass before inserting the final text. I can add names to a dictionary, choose a different key or microphone, switch between pasting and typing, and turn the history off. When history is on, the text stays on this PC and is kept for seven days by default. The recorded audio is held in memory for transcription and discarded.

Then I made it ridiculous. The interface borrows Bruce Lee's yellow and black. Nunchucks move with my voice. There are animations and sound effects. A bubble tells me to "BE WATER" while I talk. None of that improves the speech model. It does make the tool feel like mine, which is why I keep using it.

The short demo shows the important part: I talk into one app, and the words appear in another. That is the whole promise of dictation. The decorations are a bonus.

Is SaaS cooked?

For some narrow, personal workflows, it has a real problem. A generic tool may be excellent and still ask me to accept its shortcuts, its cloud connection and its monthly price. When I can build a version fitted to one person in an hour, that trade-off looks different. If you are comparing broader products rather than building one, start with what AI tools are useful for busy professionals in Hong Kong.

That does not make every SaaS company obsolete. Teams pay for reliability, support, collaboration, security controls and maintenance. My app is a personal build. I am responsible when a Windows update, a microphone driver or a new machine breaks it. Local processing also shifts setup and computing costs to me. A firm evaluating transcription for its staff would need to answer harder questions about deployment and support.

The useful lesson is smaller and more practical. Before you buy another seat, look at the one job you want the software to do. If it is repeatable and specific to you, building a small tool may be cheaper than bending your week around another subscription. The free prompt library is one place to start testing a narrow workflow. I help professionals find those jobs inside their own work. Sometimes the right answer is a product. Sometimes it is something only they would think to build.