Improving Speech Enhancement Performance by Leveraging Contextual Broad Phonetic Class Information

Previous studies have confirmed that by augmenting acoustic features with the place/manner of articulatory features, speech enhancement (SE) process can be guided to consider broad phonetic properties input when performing attain performance improvements. In this paper, we explore contextual information attributes as additional further benefit SE. More specifically, propose improve SE leveraging losses from an end-to-end automatic recognition (E2E-ASR) model predicts sequence classes (BPCs). We also developed multi-objective training ASR and perceptual train system based on a BPC-based E2E-ASR. Experimental results denoising, dereverberation, impaired tasks BPC improves performance. Moreover, trained E2E-ASR outperforms phoneme-based The suggest objectives misclassification phonemes may lead imperfect feedback, could potentially better choice. Finally, it is noted combining most-confusable targets into same calculating objective effectively