Text-conditioned Transformer for automatic pronunciation error detection

Automatic pronunciation error detection (APED) plays an important role in the domain of language learning. As for previous ASR-based APED methods, decoded results need to be aligned with target text so that errors can found out. However, since decoding process and alignment are independent, prior knowledge about is not fully utilized. In this paper, we propose use as extra condition Transformer backbone handle task. The proposed method output states consideration relationship between input speech a end-to-end fashion. Meanwhile, used decoder input, works feed-forward manner instead autoregressive inference stage, which significantly boost speed actual deployment. We set baseline model conduct several experiments on L2-Arctic dataset. demonstrate our approach obtain 8.4% relative improvement F1 score metric.