{"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/tf-locoformer-transformer-with-local-modeling","title":"TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation and Enhancement","arxiv_id":"2408.03440","date":"2024-08-06","proceeding":null,"authors":["Kohei Saijo","Gordon Wichern","François G. Germain","Zexu Pan","Jonathan Le Roux"],"abstract":"Time-frequency (TF) domain dual-path models achieve high-fidelity speech separation. While some previous state-of-the-art (SoTA) models rely on RNNs, this reliance means they lack the parallelizability, scalability, and versatility of Transformer blocks. Given the wide-ranging success of pure Transformer-based architectures in other fields, in this work we focus on removing the RNN from TF-domain dual-path models, while maintaining SoTA performance. This work presents TF-Locoformer, a Transformer-based model with LOcal-modeling by COnvolution. The model uses feed-forward networks (FFNs) with convolution layers, instead of linear layers, to capture local information, letting the self-attention focus on capturing global patterns. We place two such FFNs before and after self-attention to enhance the local-modeling capability. We also introduce a novel normalization for TF-domain dual-path models. Experiments on separation and enhancement datasets show that the proposed model meets or exceeds SoTA in multiple benchmarks with an RNN-free architecture.","url_abs":"https://arxiv.org/abs/2408.03440v1","url_pdf":"https://arxiv.org/pdf/2408.03440v1.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":"tf-locoformer-transformer-with-local-modeling","repo_url":"https://github.com/merlresearch/tf-locoformer","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[{"task_slug":"speech-enhancement","task_name":"Speech Enhancement"},{"task_slug":"speech-separation","task_name":"Speech Separation"}],"methods":[{"method_slug":"absolute-position-encodings","method_name":"Absolute Position Encodings"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"attention","method_name":"Attention"},{"method_slug":"bpe","method_name":"BPE"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"focus","method_name":"Focus"},{"method_slug":"label-smoothing","method_name":"Label Smoothing"},{"method_slug":"layer-normalization","method_name":"Layer Normalization"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"position-wise-feed-forward-layer","method_name":"Position-Wise Feed-Forward Layer"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"transformer","method_name":"Transformer"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/speech-enhancement-on-deep-noise-suppression","task":"Speech Enhancement","dataset":"Deep Noise Suppression (DNS) Challenge","model":"TF-Locoformer (M)","rank_in_archive_order":2,"of":36,"metrics":{"FLOPS (G)":"497.24","Number of parameters (M)":"15","PESQ-WB":"3.72","SI-SDR-WB":"23.3","STOI":"98.8"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-libri2mix","task":"Speech Separation","dataset":"Libri2Mix","model":"TF-Locoformer (M)","rank_in_archive_order":2,"of":10,"metrics":{"Number of parameters (M)":"15","SDRi":"22.2","SI-SDRi":"22.1"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-whamr","task":"Speech Separation","dataset":"WHAMR!","model":"TF-Locoformer (M)","rank_in_archive_order":1,"of":18,"metrics":{"Number of parameters (M)":"15","SDRi":"16.9","SI-SDRi":"18.5"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-whamr","task":"Speech Separation","dataset":"WHAMR!","model":"TF-Locoformer (S)","rank_in_archive_order":2,"of":18,"metrics":{"Number of parameters (M)":"5","SDRi":"15.9","SI-SDRi":"17.4"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (L) + DM","rank_in_archive_order":1,"of":40,"metrics":{"Number of parameters (M)":"22.5","SDRi":"25.2","SI-SDRi":"25.1"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (M) + DM","rank_in_archive_order":3,"of":40,"metrics":{"Number of parameters (M)":"15.0","SDRi":"24.7","SI-SDRi":"24.6"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (L)","rank_in_archive_order":4,"of":40,"metrics":{"Number of parameters (M)":"22.5","SDRi":"24.3","SI-SDRi":"24.2"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (M)","rank_in_archive_order":8,"of":40,"metrics":{"Number of parameters (M)":"15.0","SDRi":"23.8","SI-SDRi":"23.6"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (S) + DM","rank_in_archive_order":9,"of":40,"metrics":{"Number of parameters (M)":"5.0","SDRi":"23","SI-SDRi":"22.8"},"uses_additional_data":false},{"leaderboard":"/sota/speech-separation-on-wsj0-2mix","task":"Speech Separation","dataset":"WSJ0-2mix","model":"TF-Locoformer (S)","rank_in_archive_order":18,"of":40,"metrics":{"Number of parameters (M)":"5.0","SDRi":"22.1","SI-SDRi":"22"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2408.03440","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}