CTC-Seeded Edit Flow Decoder slashes ASR errors in just 2 steps

Non-autoregressive automatic speech recognition (ASR) traditionally starts from random, fully masked, or fixed-length token sequences, requiring many iterations to reconstruct the transcript. This paper introduces a radically different approach: use a greedy connectionist temporal classification (CTC) hypothesis as a seed, then refine it with an acoustic-conditioned Edit Flow decoder that predicts insertion, deletion, and substitution operations in parallel. The decoder is jointly trained with a CTC model using a continuous-time discrete diffusion loss. The key innovation is that the edit space is limited to the collapsed CTC output, making the refinement efficient and variable-length.

In experiments, just two edit steps yield substantial Word Error Rate (WER) reductions. Classifier-free guidance (CFG) focuses the model on audio features, further enhancing recognition quality. The authors also constrain edit proposals using CTC confidence to improve accuracy. Ablation studies validate the design choices, and decoder pretraining combined with pretrained encoder integration provides additional performance gains. Submitted to IEEE SLT 2026, this work could make real-time ASR on edge devices much more practical without sacrificing accuracy.