Papers › TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation...

TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation and Enhancement

6 Aug 2024arXiv:2408.03440archive 2025-07-28

Kohei Saijo, Gordon Wichern, François G. Germain, Zexu Pan, Jonathan Le Roux

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.

PaperPDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

merlresearch/tf-locoformer officialpytorchApache-2.0 report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Speech EnhancementSpeech Separation

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Speech Enhancement Deep Noise Suppression (DNS) Challenge TF-Locoformer (M) FLOPS (G) 497.24 #2 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge TF-Locoformer (M) Number of parameters (M) 15 #2 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge TF-Locoformer (M) PESQ-WB 3.72 #2 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge TF-Locoformer (M) SI-SDR-WB 23.3 #2 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge TF-Locoformer (M) STOI 98.8 #2 of 36 Archive leaderboard report
Speech Separation Libri2Mix TF-Locoformer (M) Number of parameters (M) 15 #2 of 10 Archive leaderboard report
Speech Separation Libri2Mix TF-Locoformer (M) SDRi 22.2 #2 of 10 Archive leaderboard report
Speech Separation Libri2Mix TF-Locoformer (M) SI-SDRi 22.1 #2 of 10 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (M) Number of parameters (M) 15 #1 of 18 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (M) SDRi 16.9 #1 of 18 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (M) SI-SDRi 18.5 #1 of 18 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (S) Number of parameters (M) 5 #2 of 18 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (S) SDRi 15.9 #2 of 18 Archive leaderboard report
Speech Separation WHAMR! TF-Locoformer (S) SI-SDRi 17.4 #2 of 18 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) + DM Number of parameters (M) 22.5 #1 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) + DM SDRi 25.2 #1 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) + DM SI-SDRi 25.1 #1 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) + DM Number of parameters (M) 15.0 #3 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) + DM SDRi 24.7 #3 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) + DM SI-SDRi 24.6 #3 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) Number of parameters (M) 22.5 #4 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) SDRi 24.3 #4 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (L) SI-SDRi 24.2 #4 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) Number of parameters (M) 15.0 #8 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) SDRi 23.8 #8 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (M) SI-SDRi 23.6 #8 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) + DM Number of parameters (M) 5.0 #9 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) + DM SDRi 23 #9 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) + DM SI-SDRi 22.8 #9 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) Number of parameters (M) 5.0 #18 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) SDRi 22.1 #18 of 40 Archive leaderboard report
Speech Separation WSJ0-2mix TF-Locoformer (S) SI-SDRi 22 #18 of 40 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEConvolutionDense ConnectionsDropoutFocusLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections