Assembly AI, an AI startup, has built Conformer-1, a speech recognition system. Conformer sets some new record scores via a couple of distinct improvements - some technical tweaks, and also some 'audio scaling laws'.

Audio scaling laws: Scaling laws are the idea that it's possible to predict ahead of time how much data (and/or compute and/or parameters) you need to achieve a given performance level. For Conformer-1, Assembly says it applied scaling laws for the speech recognition domain and used this to figure out "that for a 300 million parameter Language model, we'd need roughly 6 billion tokens of text, which corresponds to about 625K hours of speech. The team then built a dataset of 650k hours of English audio and trained Conformer on it.