On-device CPU inference is the real flex here. Optimization probably mattered as much as modeling.

Yeh, this would be helpful for the Singlish friends of mine out there!

Local CPU inference stands out. Careful optimization likely rivaled the modeling effort.

Really cool to see accent adaptation in real time — curious about benchmarks and how well this handles messy, real Zoom calls

The real achievement here isn't just quality, it's doing it streaming with tight latency on CPU while preserving speaker identity. Most VC-style work looks great offline, then falls apart once you go real-time. Nice work getting this to hold up in streaming.

Co-founder of Krisp here. 1.5B non-native English speakers in the workforce, 4x native — yet all comms infra is optimized for native accents. We spent 3 years building listener-side, on-device accent understanding. The hard parts: no parallel training data exists, the accent space is infinite, accent is entangled with voice identity, and it runs on CPU under 250ms latency. Built in Yerevan, Armenia. Beta is live and free. Happy to go deep on the ML side.

This tackles a massive yet often overlooked gap in global communication.

A majority of professionals around the world operate in English as a second language, yet most voice technology has historically been designed with native speech patterns in mind. That imbalance creates subtle barriers in everyday conversations, from team syncs to high-stakes business calls.

Building real-time, on-device accent adaptation, without clean paired datasets, across countless speech variations, while separating pronunciation patterns from speaker identity and keeping latency ultra-low on CPU, is an extraordinary technical achievement.

Deep respect for taking on something this fundamental to inclusion and clarity in the modern workplace.

Curious whether wav2vec-style embeddings played a role in your representation learning.

Finally Krisp built it! I will understand my users from interviews better, with no cognitive load and "could you please repeat that" phrasing.

The parallel data is a problem here — you can’t crowdsource ground truth because no one can record themselves with a different accent.

Latency can destroy conversational rhythm. What’s your p95 inference time? also are there any benchmarks we can see?

Kinda wild to think accent friction is basically a tech problem. Doing this in real time on CPU sounds tough. Curious how well it holds up in messy, real calls.

This feels adjacent to voice conversion research, but with stricter latency constraints.

Nice to finally see this direction of accent conversion (that is on incoming calls) in the Krisp app. This is a very meaningful feature.

Curious whether wav2vec-style embeddings played a role in your representation learning.

Great work. Natural + clear is the combo that matters.

This is a game-changer! I remember each and every call I had with an investor and feeling shy asking "can you repeat?"... thanks krisp, you changed my life!!!

This is built for international, privacy-first teams!

I would like to use such model but only if it really preserves my voice, otherwise people would understand its not me or I have to use it all the time.

Streaming constraint under 200ms changes everything. Causal modeling in speech is brutal to get right.

How did you estimate the number of IQ points?

Accent space is effectively infinite. Generalization must rely on invariants rather than enumeration.

On-device CPU inference is the real flex here! Optimization probably mattered as much as modeling.

Streaming constraint under 200ms changes everything. Causal modeling in speech is brutal to get right.

This works weirdly well — I’m honestly amazed by how good and fast it is!

Accent space is effectively infinite. Generalization must rely on invariants rather than enumeration.

will it help the barista in Starbucks get my name right finally?

This is a huge game changer !

undefined

[dead]

[dead]

[dead]

[dead]

[dead]

[dead]

[dead]

[dead]

undefined