Dan Kokotov: Speech Recognition with AI and Humans | Lex Fridman Podcast #151

Lex Fridman talks with Dan Kokotov, VP of Engineering at Rev, about how the company built a leading speech-to-text pipeline by combining machine ASR with a two-sided marketplace of human transcribers. They discuss how Rev grew out of improving the Upwork freelancer model, the data flywheel that gives Rev high-quality labeled audio, and the gap between roughly 14% machine word error rate and ~3% human accuracy. The conversation broadens into the future of searchable indexed audio, podcasting and RSS versus exclusivity deals like Spotify's, censorship and platform responsibility, and the transition from being a programmer to managing humans. They close on favorite sci-fi, dystopian films, Stalin-era history, and the meaning of life as creation.

• Lex states the episode is NOT sponsored by the guest and that no one can buy their way onto the podcast, even though Rev later agreed to sponsor.
  00:01:34
• Rev's ASR sits at roughly 14% word error rate on their test suite, versus an estimated 2-3% human word error rate.
  00:29:29
• Kokotov calls Rev's setup a 'magical flywheel': their business model is being paid to annotate the very data that trains their model.
  00:35:43
• He admits AWS and Amazon Mechanical Turk have terrible interfaces despite huge potential, and most users only touch them via API.
  00:38:46
• Both openly criticize podcast exclusivity, with Lex saying he'd turn down Spotify's $100M Joe Rogan-style deal on principle.
  00:54:52
• Kokotov argues outrage drives engagement and companies are judged by engagement, framing the core tension of platform design.
  01:01:37
• Kokotov gives a secular answer to the meaning of life: contributing to humanity through creation and raising children.
  01:25:24

• Kokotov's mother was a programmer in Russia; he played on the equivalent of a 286 PC with floppy disks as a kid.
  00:04:42
• Rev.com handles human/end-user services while Rev.ai is the ASR sub-brand, analogous to WordPress.com vs WordPress.org.
  00:09:54
• Rev was founded by an early Upwork employee to remove the friction of choosing freelancers by standardizing verticals.
  00:10:27
• Rev started with translation (birth certificates, documents) before adding audio transcription.
  00:13:30
• Most Rev transcribers are US-based work-from-home parents, students, or people seeking flexible income away from a 9-to-5.
  00:23:17
• On good audio, a skilled human transcriber takes about 2-3x the audio length to produce a transcript; real-time is very hard.
  00:27:56
• Rev measures itself against Google, Amazon, and Microsoft ASR and claims to beat them on its own internal test set.
  00:31:33
• Captioning that once cost $1/minute is now around $1.25, while automated Temi transcription is about 25 cents/minute.
  00:37:46
• Kokotov ties Brave New World's genetic sorting to modern 'genetic sorting' via assortative marriage and social-media echo chambers.
  01:09:24
• He praises Chernobyl's creator, a comedy writer with no Russia background, for capturing the spirit down to the pet bowls.
  01:16:10