Author here. I ran this to pick models for Dabri, an offline speech-to-text app for Linux I maintain: https://github.com/AshBuk/dabri
Two things surprised me. libparakeet's default of 4 threads is half as fast as one thread per physical core, and all 24 logical threads are no faster than 4.
Author here. I ran this to pick models for Dabri, an offline speech-to-text app for Linux I maintain: https://github.com/AshBuk/dabri
Two things surprised me. libparakeet's default of 4 threads is half as fast as one thread per physical core, and all 24 logical threads are no faster than 4.
Happy to answer questions about the setup.