1 paper
Oleg Rybakov, Dmitriy Serdyuk, Chengjian Zheng
Large-scale universal speech models (USM) are already used in production. However, as the model size grows, the serving cost grows too. Serving cost of large models is dominated by…