I run a small newsroom in Brooklyn. Most nights the news is a three hour public meeting on a livestream, and the job is to pull two or three real quotes out of it, cut a clip people will actually watch, and write it up before morning. That work is slow, and the easy shortcuts are the dangerous ones. A generic AI tool will hand you a clean looking quote that nobody actually said. Print that once and you lose the one thing a local newsroom runs on, which is trust.
vidtranscript is the tool I built so that cannot happen.
01What it does
You give it a link to a civic video, a community board meeting, a debate, a press conference, or you upload footage from your own camera. It transcribes the whole thing, works out who is speaking, and lets you put a name to each voice yourself. It reads the transcript and proposes the moments worth covering. You review each one, hear it played back, and approve the cuts you want to keep. Then it renders a captioned vertical clip you can post, and drafts an article to go with it. The first useful thing, a searchable transcript with names attached, is ready while you are still reading.
The model reads and suggests. It never decides what is true.
02Why it is different
The part that matters is what the language model is allowed to do, and what it is not. The model is good at reading a long transcript and pointing at the parts that are newsworthy. It is not trusted to decide what was said, what is fair, or what is allowed to ship. Three checks handle those, and each one runs in code rather than in a prompt. Each one can refuse.
When the model proposes a quote, the tool matches that text word for word against the actual words in the recording. A quote is stored as a range of word positions, and the text you read is rebuilt from those words every time. If the model writes a word that nobody spoke, the quote is rejected. There is no path for a published quote to contain a word that is not in the recording. That is mechanical, not a promise. It guarantees the quote matches the transcript. It does not claim the transcript is a perfect record of the audio, because no transcription is, which is why a person still approves every quote by ear.
The approve button stays locked until the clip has played all the way through. You cannot ship a cut you have not heard. The rule lives on the server, so a clever front end cannot get around it.
Before a debate edit can render, the tool recounts how much speaking time each candidate got in the final cut, with the moderator left out of the math. If the edit is lopsided, it will not render until you write down why. During an election season that is the difference between an edit you can defend and one you cannot. One honest limit: where people talk over each other, overlapping words get attributed to one speaker, so the count can undercount crosstalk.
03What it costs and where it runs
It runs on a laptop. Rendering the video costs nothing, because it uses ffmpeg instead of a paid video service. The only money spent is on transcription and on the model that reads the transcript, and every cent of it is written to a ledger you can query. A long meeting lands somewhere between a few cents and under a dollar, depending on how much of it is worth transcribing. The same source is never billed twice.
04What it is, and what it is not yet
I built it for Bushwick Daily, the newsroom I have run for eight years, and it is in use there now. I would rather be straight about its edges than oversell it. It is built for one operator at a time, with no accounts or hosting yet. A few pieces, like an automatic recap reel, work from the command line but are not in the web app. The core path, from a pasted link to a finished captioned clip with a verified quote, is built and runs every day.
05See it
There are two ways to go deeper. If you run a newsroom or want to test it, the product walkthrough shows the whole flow on real footage. If you are an investor, the one page pitch covers the problem, the moats, the unit economics, and the ask.
Want a walkthrough on your own footage? Email alec.meeker@gmail.com.