A deep speech-recognition network must learn from utterances of varying length. Starting with the harder cases can make early training unstable. Curriculum learning changes the order of examples so the network encounters easier cases before more challenging ones. section 4.4.2
SortaGrad in Deep Speech 2
SortaGrad sorted training sequences by length during the first epoch. Shorter utterances, treated as typically easier, came first. This helped reduce numerical instability and supported reliable early convergence. Alongside sequence-wise batch normalization, it helped Deep Speech 2 train deeper networks and improve transcription accuracy and stability. section 4.4.2
This was one part of a larger engineering effort, not a stand-alone explanation for the system’s gains. Deep Speech 2 combined curriculum learning with architectural refinements, computational optimizations, and normalization—a pragmatic approach to improving end-to-end learning. section 4.4.5
