TechBeret

@techberet
27 Followers
12 Following
11 Posts
Software and firmware engineer in the Pacific Northwest.
Bloghttps://techberet.com
Show threadTechBeretSep 7, 2024
@chadaustin Unfortunately very variable. I have a couple keyboards from eBay which look brand new, and others that look very worn out. The image quality from many eBay sellers also don’t make it easy to judge.
Show threadTechBeretSep 7, 2024
@chadaustin it is actually pretty easy and cheap to find dongle-less keyboards on eBay. It was how I stocked up for my project (5 spare keyboards), as I didn’t want to open any of the 3 unopened ones I had stockpiled.
Show threadTechBeretSep 7, 2024
Special thanks to @chadaustin who had a great write up of his Sculpt PCB project which was extremely helpful in getting mine kicked off: https://chadaustin.me/2021/02/wired-sculpt/
Microsoft Sculpt Wired Conversion Mod

I made a control board for the Microsoft Sculpt wireless keyboard that converts it to wired USB, and now my favorite keyboard is even better.

Chad Austin
TechBeretSep 7, 2024
Wrote up a long blog post about my latest project: designing a custom PCB for the Microsoft Sculpt Ergonomic keyboard, which stopped being manufactured causing prices to skyrocket. Covered topics include processor selection, proof of concept work, KiCad (including some useful tips), USB-C port wiring, a brief mention of OpenSCAD, assembly, and testing. Full source code for both HW and SW is included. https://techberet.com/2024/09/07/replacement-pcb-for-sculpt-rp2040
Designing a Replacement PCB for the Microsoft Sculpt Ergonomic Keyboard - TechBeret Blog

Table of Contents (apologies, this is a long post)

Show threadTechBeretFeb 9, 2023
@leonzandman @marcoarment @siracusa @caseyliss Benchmarks run against GPU were run using standard Python code.
Show threadTechBeretFeb 9, 2023
@leonzandman @Modug @marcoarment @caseyliss The stats page already has that info collected, and even has an interactive graph looking at word spoken per episode. It’s an interesting ratio of roughly 3:2:1 of John:Marco:Casey words spoken across all episodes. https://catatp.fm/statistics/
Statistics · catatp.fm

TechBeretFeb 7, 2023
Happy to announce that catatp.fm is now fully transitioned over to Whisper for speech transcription. The entire backlog has been processed using Whisper’s large v2 model to maximize accuracy. Wrote up a blog post about learnings from the transition, how I handled Whisper’s often inaccurate timestamps, and some fun speed benchmarks comparing some different systems for transcription, and a test comparing Python Whisper to whisper.cpp. https://techberet.com/2023/02/07/catatpfm-now-with-whisper
@marcoarment @siracusa @caseyliss
catatp.fm - Now with Whisper Powered Transcription - TechBeret Blog

Another Leap in Transcription Accuracy

Show threadTechBeretJan 21, 2023
@_Davidsmith FYI if you can stomach the speed hit Whisper's large-v2 model does a better job:
Show threadTechBeretJan 16, 2023
@marcoshuerta no need to currently, unfortunately the whisper transition is in my half completed side projects category and I don’t have the time today to dive back into it. I’m not the only one running into the issue though, as evidenced by this GitHub PR: https://github.com/lhotse-speech/lhotse/pull/928
Remove negative duration segments from whisper by desh2608 · Pull Request #928 · lhotse-speech/lhotse

This PR addresses #891. Remove segments with non-positive duration from the whisper output Segment post-processing to force non-overlapping is made optional (disabled by default) Allow overlapping...

GitHub
Show threadTechBeretJan 15, 2023
@_Davidsmith @siracusa Pyannote is what I used for catatp.fm speaker identification, but they recently changed their model which dramatically reduced accuracy on long form audio clips (like podcasts). I was able to revert back manually and keep using their old model, but in the future I plan on investigating nVidia NeMo, which I've seen anecdotal reports of high accuracy on.