Introducing Atlas 1. Willow's new frontier speech-to-text model. It outperforms ElevenLabs, Deepgram, OpenAI, and more by a wide margin. Built on the first scalable, human-powered transcription infrastructure ever built for real-time dictation.

Most models score 5-7% word error rate on clean audio. In real-world conditions they fall to 10-15%. Atlas 1 holds at 1.2% on clean audio and 2.1% in production. The gap widens in noisy environments.

Atlas 1 rolls out to all Willow users today. Get access:

These guys dictate to their devices so much now that they never need to remove their hands from their pockets!

@ChrisWordbox True!

tuff

@sollnikhil

Just tested it in Arabic. It has no idea what I'm saying. Selected English and Arabic as the language -> failed (typed gibberish in English) Arabic only -> failed Auto detect language -> failed Uninstalled after 10 minutes.

Hi Abdullah, that's strange to hear. Can you please double check your microphone settings and make sure you're set as the built-in microphone? That's typically the most common reason for why you see typed gibberish rather than completed dictations. Please send an email to [email protected] for direct support!

Plans to provide APIs for other use cases?

Soon! We're at capacity right now

the hands in the pockets LOL

@sama

Holy smokes, just downloaded and it took me 30-seconds to switch from Wispr Flow. This can't be real.

Let's go!

beating giants like OpenAI and Eleven Labs is a huge feat

it's time for AI companions to thrive

This is what happens when the bottleneck shifts from models to data quality. Human-powered infrastructure at scale is a serious unlock.

This is the future.

Without any doubts.

brb, switching from wispr

🫡

How is this even possible 👀

947 specialists. 14 months. $8.4M in operational infrastructure.

Crazy

This could be one of the most uncomfortable launch videos I’ve seen

absolutely wild boys

Sat up immediately I saw "it outperforms ElevenLabs"

@PM_Mide Not only do we outperform, we leapfrogged the entire industry 5X.

Huge W for Willow

API?

repeat after @HanchungLee: the model is the product

Feels like the real advantage here is system design, not just the model. Curious how much of the gain comes from infra vs model improvements?

humans as a service. unreal work

Okay guys we will try it. Please don’t beat us

why all hands in pockets

The improvement in noisy settings is especially interesting here.

🫡

Why are they all wearing dark shirts

@tiffanyyan_x

how does it compare with parakeet v3 and cohere-transcribe in terms of RTF?

947 specialists. $8.4M. 14 months. This is the data infrastructure story nobody in AI is talking about. Everyone focuses on who has the biggest model. The real moat is who has the highest quality human-annotated training data. "Human-powered infrastructure at scale is a serious unlock" — this is the quiet revolution in AI rn.

> looks like a fun place to work at

Dumb question, how does it compare to @WisprFlow in day to day use ?

@WisprFlow We're so accurate that it almost feels like someone is actually listening to you...

You got APIs ??

Soon, once we hire more transcription specialists!

Where can I apply? And what are the qualifications

@ycombinator Happy April Fools

Hot take: if your voice AI still makes errors, you’re using the wrong one. W Willow 🙌

let's go

SF is truly not a real place

Willow is the zeitgeist.

Seems that way. Excited to try

Human-powered?

@WillowVoiceAI will this support cantonese?

We currently support Cantonese!

no way this is real