Mohan Xu^*, Kai Li^*, Guo Chen, Xiaolin Hu
Tsinghua University, Beijing, China
^*Equal contribution
📜 ICLR 2025 | 🎶 Demo | 🤗 Dataset

In this paper, we propose a speech separation model with significantly reduced parameter size and computational cost: Time-Frequency Interleaved Gain Extraction and Reconstruction Network (TIGER). TIGER leverages prior knowledge to divide frequency bands and applies compression on frequency information. We employ a multi-scale selective attention (MSA) module to extract contextual features, while introducing a full-frequency-frame attention (F^3A) module to capture both temporal and frequency contextual information. Additionally, to more realistically evaluate the performance of speech separation models in complex acoustic environments, we introduce a novel dataset called EchoSet. This dataset includes noise and more realistic reverberation (e.g., considering object occlusions and material properties), with speech from two speakers overlapping at random proportions. Experimental results demonstrated that TIGER significantly outperformed state-of-the-art (SOTA) model TF-GridNet on the EchoSet dataset in both inference speed and separation quality, while reducing the number of parameters by 94.3% and the MACs by 95.3%. These results indicate that by utilizing frequency band-split and interleaved modeling structures, TIGER achieves a substantial reduction in parameters and computational costs while maintaining high performance. Notably, TIGER is the first speech separation model with fewer than 1 million parameters that achieves performance close to the SOTA model.

Overall pipeline of the model architecture of TIGER and its modules.

Performance comparisons of TIGER and other existing separation models on Libri2Mix, LRS2-2Mix, and EchoSet. Bold indicates optimal performance, and italics indicate suboptimal performance.

Efficiency comparisons of TIGER and other models.

Comparison of performance and efficiency of cinematic sound separation models on DnR. '*' means the result comes from the original paper of DnR.

If you have any questions, please feel free to contact us via tsinghua.kaili@gmail.com.