2 years ago, we wanted to create a better way to learn pronunciation that was fun, cheap, and effective. Fast forward, Koel has taken our phonetic model to an accessible consumer app that can give you instant pronunciation feedback. We demoed @spc, check out the Slayspeech app!
an AI in the top 1% of super smash bros melee, a 7MB model that teaches accents, visualized math behind stable pre-training, and more.
november demo night at @spc showed how much good research can look like play.
(0:28) @jeffreycider & @jacknzheng: why models don't blow up in pre-training
(8:37) @KoelLabs: an accent coach and a 7MB open-source phonetic model
(15:17) @ericyuegu: pre-training a transformer on Super Smash Bros Melee
We studied 16 systems released from 2021-2025 and found that newer speech models don't improve for everyone. Accent gaps didn't reliably shrink. Some minority accents saw large regressions while mainstream US English remained relatively stable: https://t.co/zu4EQbkEgQ
Training a new model and deciding to spend your $$ on more data or higher quality data? Our Interspeech study shows that for our task, there was a clear threshold after which only high quality quality data was worth it. The threshold was surprisingly low: https://t.co/AUBNoYE2Vo
Super cool. We tried this on our speechโIPA models. Frontier LLMs couldn't improve our best, but pushed our 7.4MB model (distilled with custom architectures + QAT) from #8 to #4; behind models 100x larger.
Could LLMs soon generate tiny models on demand?
https://t.co/AornYvZW71
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! ๐งต
Blog: https://t.co/bOWCcFhCVq