{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/perceptual-loss-based-speech-denoising-with","title":"Perceptual Loss based Speech Denoising with an ensemble of Audio Pattern Recognition and Self-Supervised Models","arxiv_id":"2010.11860","date":"2020-10-22","proceeding":null,"authors":[],"abstract":"Deep learning based speech denoising still suffers from the challenge of\nimproving perceptual quality of enhanced signals. We introduce a generalized\nframework called Perceptual Ensemble Regularization Loss (PERL) built on the\nidea of perceptual losses. Perceptual loss discourages distortion to certain\nspeech properties and we analyze it using six large-scale pre-trained models:\nspeaker classification, acoustic model, speaker embedding, emotion\nclassification, and two self-supervised speech encoders (PASE+, wav2vec 2.0).\nWe first build a strong baseline (w/o PERL) using Conformer Transformer\nNetworks on the popular enhancement benchmark called VCTK-DEMAND. Using\nauxiliary models one at a time, we find acoustic event and self-supervised\nmodel PASE+ to be most effective. Our best model (PERL-AE) only uses acoustic\nevent model (utilizing AudioSet) to outperform state-of-the-art methods on\nmajor perceptual metrics. To explore if denoising can leverage full framework,\nwe use all networks but find that our seven-loss formulation suffers from the\nchallenges of Multi-Task Learning. Finally, we report a critical observation\nthat state-of-the-art Multi-Task weight learning methods cannot outperform hand\ntuning, perhaps due to challenges of domain mismatch and weak complementarity\nof losses.","url_abs":"http://arxiv.org/abs/2010.11860v1","url_pdf":"http://arxiv.org/pdf/2010.11860v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"perceptual-loss-based-speech-denoising-with","repo_url":"https://github.com/saurabh-kataria/PERL-samples","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":null}],"tasks":[{"task_slug":"denoising","task_name":"Denoising"},{"task_slug":"emotion-classification","task_name":"Emotion Classification"},{"task_slug":"multi-task-learning","task_name":"Multi-Task Learning"},{"task_slug":"speech-denoising","task_name":"Speech Denoising"}],"methods":[{"method_slug":"pase","method_name":"PASE+"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/speech-enhancement-on-demand","task":"Speech Enhancement","dataset":"VoiceBank + DEMAND","model":"PERL-AE","rank_in_archive_order":24,"of":42,"metrics":{"CBAK":"3.53","COVL":"3.83","CSIG":"4.43","PESQ (wb)":"3.17"},"uses_additional_data":true}],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}