Eman Khoubian

Founder. Operator. Investor.
@getwhisperai

[Eman Khoubian @emankhoubian]

@noahkagan Congrats! What are you using as the transcription engine?

[Eman Khoubian @emankhoubian]

@amasad WhisperAI just processed ~145,000 hours of audio and video in October.

And it was all done on @Replit

[Eman Khoubian @emankhoubian]

@markfenner @Replit Thank you 🙏

[Eman Khoubian @emankhoubian]

I built WhisperAI as a side project using @Replit.

16 months later, we just processed 145,000 hours of audio and video in a single month.

What started as a side project has quickly grown into one of the world’s leading platforms for audio and video transcription.

And we’re just getting started.

[Eman Khoubian @emankhoubian]

@Bagdex @Replit Thank you 🙏

[Eman Khoubian @emankhoubian]

https://t.co/0N2UbZoth1

[Eman Khoubian @emankhoubian]

@Austen Try WhisperAI

https://t.co/KErhWHLcQG

[Eman Khoubian @emankhoubian]

January 2026: ~22,000 hours transcribed
October 2026: ~145,000 hours

That’s almost 7x in 10 months

[Eman Khoubian @emankhoubian]

WhisperAI just had a record month in October: 145,000 hours of audio and video transcribed.
That’s 40% more than Trint’s 2025 monthly average.

[Eman Khoubian @emankhoubian]

@AarushYadav47 How does it handle different languages? Background noise?
Curious what happens when it’s live vs in testing.

[Eman Khoubian @emankhoubian]

@Jason @meetgranola @WisprFlow WhisperAI does it all and more. Give it a try

[Eman Khoubian @emankhoubian]

@kwindla The foundation is the part to watch here. Pipecat has more than 200 integrations and almost 400 contributors, so vendor neutral only holds if the steering committee is not just Deepgram people.

[Eman Khoubian @emankhoubian]

People change their mind inside a single call, and a transcript holds both versions flat.

Someone opens with a hard no. Forty minutes later, after the numbers get walked through, they agree to a smaller version with a condition attached. Both statements sit in the file with the same weight, the same formatting, nothing marking one of them as the position that replaced the other.

Search makes it worse. The first mention of the topic is the one a search finds fastest, and the first mention is usually the opening position, stated before anyone had heard the full picture. A reader who was not on the call searches the file, lands on the early no, and quotes it as the decision.

In the room, nobody makes that mistake. Everyone heard the order things happened in, and the later statement carried the earlier one away with it. On the page, order is the only signal left, and it is a weak one. Timestamps show sequence, but nothing in a block of text says this line cancels the one twenty pages up.

When a decision matters, I read a transcript backwards from the end. The last time a topic comes up is usually where it settled. The first time is where it started, and those two are confused with each other more often than any misheard word costs anyone.

[Eman Khoubian @emankhoubian]

@meetgranola Unlimited words for free is the number that matters here. No meter running changes how long people are willing to talk, and Windows and iOS still coming soon means the phone test for that has not happened yet.

[Eman Khoubian @emankhoubian]

@raycast Dictation inside the primary keyboard removes the app switch that kills most voice typing. I would watch week two usage on short replies, where swipe typing usually wins people back.

[Eman Khoubian @emankhoubian]

Wind is the noise a transcript shows you as silence, then as wrong words.

A person dictating outside can sound clear to anyone standing next to them. The microphone gets a different recording. A gust hits the mic in a way a nearby listener never hears, and for a second or two the voice drops under a low roar that leaves no word in the transcript at all.

The damage is uneven. Short words go first. A no, a not, a date said quickly while walking. Those small words change the meaning of the sentence around them, and wind covers them most easily because they take the least time to say.

I judge outdoor recordings by their worst gust, not by their average clarity. Play the sections where the speaker turned a corner or crossed an open space. If the small words survived there, the rest of the file will usually hold up. If they vanished, the clean sections in between give you a false sense of how much you can trust the whole note.

A foam cover helps more than a better model does here. So does turning your back to the wind, or waiting ten seconds to say the important part. The model can only work with what reached the microphone. Once a gust takes a word, there is nothing left in the audio to recover.

[Eman Khoubian @emankhoubian]

@cartesia @ArtificialAnlys @ServiceNow $0.03/min is the number I would watch more than the top ranking on both benches. The messy 15 minute call where turn-taking keeps breaking costs more than a bench average shows, and that is the call that eats the margin.

[Eman Khoubian @emankhoubian]

@moinnadeem @phonic_co @firstround @Lux_Capital @felicis The 44-minute call with 123 tool calls is the part I would trust more than the 87.1 index score. Most voice demos fall apart around minute ten when the caller changes their mind mid-task and the tool calls start piling up.

[Eman Khoubian @emankhoubian]

The room is in every recording, and a transcript never shows it.

An empty room with hard walls adds echo to every word. The voice reaches the microphone twice, once direct and once bounced off the wall a fraction of a second later. In the room your brain sorts that out without trying. In a transcript the two arrivals blur together, and the words that suffer first are the short ones with hard edges, the consonants that separate one word from the next.

That is why the same person can sound clear in a furnished office and muddy in an empty conference room on the same laptop. Nothing about the speaker changed. Nothing about the model changed. The room changed, and the room is the one part of the setup nobody thinks to describe when a transcript comes back rough.

I check the room before I check anything else on a bad file. Soft rooms forgive a cheap microphone. Hard, empty rooms punish a good one. A few soft surfaces near the speaker do more for the transcript than moving the microphone six inches closer, because they kill the second copy of the voice before it reaches the mic.

A transcript shows you words. It never shows you the room they were spoken in, even though the room helped decide which words survived.