Hazz Wispr

Problem
I dictate a lot of my work, from client messages to prompts for AI tools, and I was paying a monthly subscription for Wispr Flow to do it. Paid tools like that process your voice in the cloud, and for client work I’d rather nothing left my laptop. The free option, macOS dictation, isn’t smart enough. It types every “um”, can’t cope when I correct myself mid-sentence, and gets names and jargon wrong. I wanted the paid experience for £0 a month, running entirely on my own Mac.
Solution
I designed and built Hazz Wispr, a Mac menu bar app, using Claude Code as my engineering partner. You hold Right Option, speak, and let go, and the text appears wherever your cursor is, in any app.
Speech is transcribed on the Mac itself with OpenAI’s Whisper model. A local AI model then tidies the text. It removes filler words and keeps only the corrected version when I change my mind (“Tuesday, no wait, Wednesday” becomes “Wednesday”). It also fixes misheard words from context, so “BB to B” comes out as “B2B”. A personal dictionary makes sure brand names, like my agency Leunam, are spelled right first time. A small dashboard shows my stats and history.
Privacy shaped the whole build. Audio and text never leave the device, the dashboard only works on my own Mac, and a guard rejects any AI output that rewords what I actually said. It’s a proper signed Mac app that starts at login, and it’s open source on GitHub.
Tech: Python, MLX, Whisper, Ollama (Qwen 2.5), SQLite, vanilla JS and a small C launcher. About 2,000 lines of code.
Result
Hazz Wispr replaced a paid monthly subscription for nothing. In the first three weeks I used it for 403 dictations and 9,207 words, saving about two and a half hours compared with typing. It works fully offline, and text appears about 1–2 seconds after I stop talking.
Getting there taught me something. My first attempt at speeding it up made it faster but noticeably less accurate. I spotted the drop in my own usage data, traced the cause and rebuilt it so speed never comes at the cost of accuracy.
The code is on GitHub: github.com/harrymanuel/hazz-wispr-dictation ↗


