Controlled speech experiments: joint training from scratch always beats separate pretraining
giffmana · x · 2026-11-16
Albert Zeyer responds to giffmana on whether speech models need separate pretraining, sharing experimental findings.
- The team ran multiple controlled comparisons for speech (data-controlled, not compute-controlled, all trained to convergence).
- Result: joint training from scratch was always better.
- Caveat: training criteria and other factors also matter, so pretraining isn't the only lever.
More from Research
- Terence Tao: math has moved from proof scarcity to proof abundance — haider1 · 2026-11-16
- Wrong AI advice hurts doctors' judgment more than right advice helps: study of 16,641 decisions — ValerioCapraro · 2026-11-16
- Free online book 'The Science of Computing' links PDEs, numerics, and cache performance — burkov · 2026-11-16
- Free Textbook: Algebraic Topology for Data Scientists and TDA — burkov · 2026-11-16
- CVPR'27 Paper Deadline in 2 Days as Major AI Conference DDLs Cluster — AlexTensor · 2026-11-16
- Meta & NYU paper: vision prior unnecessary, dedicated capacity placed early is what matters — burny_tech · 2026-11-16