Wispr Flow makes everyday life considerably easier because I can dictate everything straight into the computer, which makes everyday tasks like writing emails or typing up neurological findings significantly faster.
At the same time, Wispr Flow really annoys me with its persistent pop-up dialogs that refuse to go away for many seconds. Guys – are you using your own software? Whoever would enjoy that! It also struggles heavily with specialized medical terminology.
By now, I have manually typed in around 150 medical terms. So it’s slowly getting there.
Parakeet from NVIDIA a faster dictation model
this is what I have heared in a youtube: Parakeet outruns Wispr Flow 10 x in speed.
So, using my AI, I had her a Wispr Flow clone built, along with a custom medical dictionary containing 1,500 terms.
It works like a charm; I’ve already set it up across all the computers in the clinic.
Based on NVIDIA’s Parakeet, here is a 1:1 Wispr Flow clone that uses a local model—meaning it processes everything on the laptop itself and uploads nothing to the cloud:
On the first launch, it needs to download an AI model of roughly 500 MB to handle speech recognition—it is said to support 20 European languages, I can only try German and English and they work.
If you’d like the source code, feel free to contact me.
ChatGPT Audio – better, faster, more reliable
If you talk with ChatGPT – e.g. during a car ride exploring e.g. the trajectomes of the Nucleus accumbens – you are simply impressed of the speed and understanding capabilities!
So I got another dictation / transcription program based on OpenAI’s chat function.
For my workflow, this one is now completely spot on.
I got it pre-loaded with 1,500 medical terms, and unlike Wispr Flow, it easily recognizes dystonia, dyskinesia, and dysplasia.
Audio-Transcriber (c) by HeliR + Codex
However, my Audio-Transcriber relies on a cloud-based model, meaning spoken audio does not remain strictly on the computer.
While above Parakeet-Flow keeps everything local, the OpenAI Audio Transcriber requires an active internet connection—a factor that may matter when dealing with sensitive information.
To run it, you will need an OpenAI API key—in other words, an OpenAI platform account where you can generate an API key and add account credit. Yesterday, for instance, I spent about 10 cents dictating, which is quite reasonable. The above Parakeet-based program does not charge you anything.
The app itself is self-explanatory and includes two display modes: a compact live-transcription window (see below) where you can watch every word trickle in, or a minimal audio bar simply showing that recording is active.
I do not translate the main window to english, its so easy and self explanatory, in case you need to understand the text: you can copy the picture and get it translated via your phone or any AI. If you like I can provide you with the source to develop it further.
English and German are recognized and transcribed equally well.
Wispr Flow messed wrongly
Had they not annoyed me so thoroughly with constant promotional pop-ups, I would have happily and gratefully continued paying for Wispr Flow. In the end its a wonderful tool – just not for doctors – why they do not offer automatic word-sets I do not understand, as I order Codex “get 1500 of the most used medical terms as additional vocabulary into the program” and 3 minutes later its done.
Now I have my own audio transcriber: it runs silently, behaves cleanly, and is both faster and far superior. Thanks, Codex!
Enjoy using it!
final considerations
that a 63 year old fart as me is able to build such tools in about 30 minutes in the evening is kind of unimaginable – to me as a guy who programmed 40y ago in Assembler and later on in Turbo-Pascal and VBA more than 150.000 lines of code in my live!
SaaS (Software as a Service) is essentially dead.
Yesterday I watched a YouTube video where someone recreated Calendly—an online appointment booking tool—across three different AIs (Kimi k1.5, Codex, Claude Code) just to compare their abilities and giving notes mainly on the look of the frontend, because – of course the functionality was working.
Calendly is valued at $2 billion and basically dead. Probably not for much longer!
Google CENSORSHIP!
Google censors my homepage quite a bit, sometimes I am not even able to find my articles on Google. So please sign up for the newsletter and share it with friends or via Facebook and use the search function on my website. Follow me on Twitter, where I also announce important articles.
