TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation

Researchers have developed TIGER, a new speech separation model that reduces parameters by 94.3% and computational costs by 95.3% while outperforming current state-of-the-art models. The team also introduced EchoSet, a new dataset with realistic acoustic environments that shows better generalization for speech separation models.

• →TIGER achieves superior performance compared to state-of-the-art TF-GridNet while using 94.3% fewer parameters and 95.3% fewer MACs.
• →The model uses time-frequency interleaved gain extraction and multi-scale selective attention for efficient speech separation.
• →EchoSet dataset includes realistic reverberation effects considering object occlusions and material properties for better model evaluation.
• →Models trained on EchoSet demonstrated better generalization ability compared to other datasets when tested on real-world data.
• →The research addresses the critical need for high efficiency in low-latency speech processing systems.