YesPress Dossier · Speech-to-Text
Rev.
The company that decided the smartest way to teach a machine to listen was to pay people to do it first.
The company that decided the smartest way to teach a machine to listen was to pay people to do it first.
Right now, a deposition recorded in a fluorescent-lit room, a podcast taped in a closet, and a university lecture half-mumbled into a laptop mic are all being turned into clean, searchable text. The thing they have in common is Rev. It is the plumbing under the spoken word - rarely visible, almost always working.
Rev sells one deceptively simple promise: give it audio, get back words you can trust. Captions for the deaf and hard of hearing. Subtitles in another language. A legal transcript a court will accept. The catch is that "trust" is the hardest word in the sentence. Anyone can produce a transcript. Producing one that is right - on accents, crosstalk, legal jargon, and the occasional cough - is the entire business.
"Anyone can produce a transcript. Producing one that is right is the entire business."YesPress · the thesis in one line
Today Rev runs both halves of that bet at once. A proprietary AI does the fast, cheap pass. A global network of human transcriptionists does the part the machine still flubs. The two have learned to live together, which is more than can be said for most of the AI industry.
For most of human history, if you said something out loud, it was gone. You could record it, but a recording is a locked box. You cannot search it, skim it, quote it, or hand it to a judge. An hour of audio is an hour of your life to get the one quote you need.
The obvious fix - have a computer do it - kept colliding with reality. Early speech recognition was confident and wrong, which is the worst combination. It heard "wreck a nice beach" when you said "recognize speech." Good enough for dictating a text message. Nowhere near good enough for a courtroom, a broadcast caption, or a researcher quoting a subject verbatim.
"Early speech recognition was confident and wrong - the worst possible combination."On why "good enough" was never good enough
So the market split. You could have fast and cheap, or accurate and human, but not both. That gap - between what machines could do and what serious work demanded - is the room Rev decided to live in.
Rev was founded in 2010 by six entrepreneurs - Jason Chicola, Paul Huck, Dan Kokotov, David Abrameto, Josh Breinlinger, and Mark Chen - whose roots traced back to MIT and to oDesk, the freelancing platform that later became Upwork. Marketplaces of remote workers were, quite literally, in their professional DNA.
Their bet was contrarian for its era. Instead of waiting for AI to get good enough, they built a two-sided marketplace: customers on one side, a distributed army of freelance transcriptionists on the other. People who wanted the freedom to work from home typed; customers got accuracy the machines couldn't yet touch. It worked years before "remote work" became a slogan.
"They didn't wait for the AI to get good. They hired people - and the AI learned from them."On the quiet logic of the marketplace
Here is the twist nobody planned out loud: every transcript a human cleaned up was a training example. Rev spent a decade paying people to produce flawless text - and accidentally built a dataset no pure-software competitor could match.
Rev is not one product so much as a toolkit for anything spoken. Drop in a file or a meeting link, choose your trade-off between speed and human-grade accuracy, and get back text that does real work. A filmmaker captions a documentary so a deaf audience can watch it. A market researcher pulls quotes from forty hours of interviews without listening to a single one. A solo podcaster turns an episode into a blog post before lunch. A law firm hands a recorded deposition to a tool that summarizes it overnight. Same engine underneath, very different lives on top of it.
Transcripts cleaned and reviewed by people, targeting up to 99% accuracy - the version a court or a journalist can stand behind.
Launched in 2018. The fast, automated pass powered by Rev's own speech engine for when minutes matter more than perfection.
Closed captions for accessibility and multilingual subtitles for reaching audiences who don't speak your language.
A developer API that drops Rev's recognition into your own app, so you can build listening into software.
Open ASR and speaker-diarization models trained on 200,000 hours of human transcripts - given away to researchers.
AI deposition summaries and evidence tools, deepened by the 2025 acquisition of SmartDepo - aimed squarely at law.
"You can dial Reverb from every 'um' and false start to clean and readable. Verbatim is now a slider."On a feature only a transcription nerd could love
Claims are cheap. Here is what's actually on the table - the customers, the scale, and the one number that explains why anyone pays for human review at all.
The last few percent - the names, the legal terms, the two people talking over each other - is where a transcript either earns trust or loses it. That gap is the whole reason Rev keeps a human in the loop.
"The last few percent isn't a rounding error. In a courtroom, it's the case."On why accuracy is binary when it matters
Rev's mission started broad: make the world's spoken words accessible, accurate, and searchable. Lately it has aimed somewhere specific - the legal system, where overworked teams drown in hours of testimony and evidence, and where a missed sentence has consequences.
The pitch to lawyers is the same pitch as always, just higher stakes: let trustworthy, human-verified AI surface the key facts faster, strip out the tedious work, and free people to do the part that actually requires a human. The company is blunt that AI in law has to be accurate, transparent, and verifiable - which, conveniently, is the one kind of AI Rev has spent fifteen years building. It is a tidy bit of corporate karma. The thing that made transcription trustworthy - never letting the machine have the last word - is exactly the thing a courtroom demands. Rev didn't pivot to legal AI so much as walk through a door its own history had already unlocked.
"Trustworthy AI isn't a feature you bolt on. For Rev it's the fifteen-year head start."On turning a dataset into a moat
Return to where we started. The deposition, the podcast, the half-mumbled lecture. A decade ago, each of those was a locked box - hours you'd have to sit through to find the one line that mattered. Now they're documents. Searchable. Quotable. Captioned for someone who couldn't hear them. Translated for someone who couldn't understand them.
That's the quiet thing Rev changed. Not by replacing the humans, and not by trusting the machines blindly, but by getting the two to do what each is good at. The spoken word stopped being a prison and started being data you can actually use.
"The spoken word stopped being a prison. Somewhere, an audio file just became a sentence."YesPress · closing the loop