Christian Vestergaard (@CVestergaard_) header
Christian Vestergaard (@CVestergaard_) profile picture

Work: Founder (YC backed). Escape: Cycling. Personality: Simple.

San Francisco
Joined October 2015
It's time for new training regimes in speech. We believe audio-native RL environments with fine-grained rubrics (naturalness x task completion) is part of the answer. We've been doing research on this and can soon release a preview of our own post-training. Truly exciting times!
3
2
17
606
Too frugal to pay for an expensive launch video - so we found a free content studio and spent $250 on a blackboard.
1
11
276
In case Super Intelligence is coming, make sure to treat your robots well.
1
1
5
173
New levels to being locked-in
10
208
Got tired of reading papers on my laptop, so I bought a printer for 40 bucks. Insane bang for buck! Most delightful onboarding UX I’ve had in a while, believe it or not.
3
5
211
Morning SF! Beauty of Fall, weather of Summer.
1
5
152
We're back as the #1 trending audio dataset on @huggingface 🤗 We held the top spot for 17 days in September, and interest has picked up again since we released our turn-taking model trained on OpenYap-1k. Soon we can reveal some very interesting work. Keep building!
5
12
372
We trained a small turn-taking model and it’s #1 on @sesame's TurnBench. We don't really believe in the future of turn-taking models, as conversation is too complex a dance to be learnt by a low-parameter endpointing mechanism. Instead, as @juberti correctly quoted the other day: "it needs to be done by a large model that has significant conversational context and strong priors about human interactions". However, we wanted to demonstrate what difference data can make. Most "natural" conversational speech data available today comes from strangers paired up on a gig platform and given a topic. Our top spot is mostly earned by training in it on real conversations between friends and family, rather than implementing clever algorithmic ideas.
6
3
7
322