Trancription is simply the process of turning human speech into text. While mostly an easy and solved problem through Artificial Intelligence, there are a few edge cases like the Cocktail Party Problem, where overlapping speakers (people tend to slightly interupt each other) make it difficult to parse speech. Humans can do it because they have learned to focus on a select sample of speakers, so even though the cocktail party is loud and filled with speech, you still understand the people you are talking to perfectly.
Transcription in My Workflow
made myself a little tool with nvidieas parakeet nemo model, using the MLX version to run on apple silicon, bound to alt+r to transcribe my speech to clipboard.
Transcription
This tool has been around for a while now. I wanted to write about it, given how much I’ve been using it recently. I have hit a bottleneck in how fast I can type. I type quite fast, and my typing speed doesn’t seem to be increasing anymore, despite me doing more monkeytype practice runs. Now, no matter how fast I could type, nothing could really be faster. then how fast people can talk. And I had seen tools like Whisper Flow that were exceptional at this ability. However, Whisper flow costs money and runs on someone else’s server. That makes no sense for such a lightweight program, so I decided to make my own. This uses the Nemo Parakeet model, a very lightweight model from Nvidia, which is great, but I’m on a MacBook Pro. So what I ended up doing was I found a library that does Nemo Parakeet on MLX architecture and created a Python interface that runs in the background so that when I press Alt R, it starts listening. When I stop, it creates a recording and then starts processing it, transcribing it, and putting up my clipboard. The new update I added allows it to start transcribing as I am talking to speed up the transcription time when I finally stop the recording. This is a good feature that has made this an incredibly fast tool and one that I can no longer live without.
Transcriptions will use the callout as show above.
03-13-2026 Update - HealthTRAC
neat thing is I now get to do a lot of work with Transcription/Diarization for my Undergraduate Research. This will be playing around a lot more with the Mathematics side of transcription. Looking at a lot of algorithms, and even lead me into better understand Hidden Markov Model implementations through Viterbi decoders and the Forward Backward Algorithm.