Fast Forced Aligner

About

Fast Forced Aligner is a Python package of CUDA kernels that takes a CTC model's log probabilities and transcript token IDs and returns a token label and score for every audio frame, batched across recordings of different lengths without padding. It includes an optimized kernel and a conventional dynamic-programming mode behind the same API. The author reports 11.5x faster than a corrected CPU baseline and 208x faster than k2 CUDA on a ten-workload suite, with possible differences on adversarial near-tie inputs.