3S-TSE: Efficient Three-Stage Target Speaker Extraction for Real-Time and Low-Resource Applications

Target speaker extraction (TSE) aims to isolate a specific voice frommultiple mixed speakers relying on a registerd sample. Since voiceprintfeatures usually vary greatly, current end-to-end neural networks require largemodel parameters which are computational intensive and impractical forreal-time applications, espetially on resource-constrained platforms. In thispaper, we address the TSE task using microphone array and introduce a novelthree-stage solution that systematically decouples the process: First, a neuralnetwork is trained to estimate the direction of the target speaker. Second,with the direction determined, the Generalized Sidelobe Canceller (GSC) is usedto extract the target speech. Third, an Inplace Convolutional Recurrent NeuralNetwork (ICRN) acts as a denoising post-processor, refining the GSC output toyield the final separated speech. Our approach delivers superior performancewhile drastically reducing computational load, setting a new standard forefficient real-time target speaker extraction.