Target Speaker Extraction

Target speaker extraction separates or enhances a selected person’s speech in mixtures, helping listeners follow conversations and improving downstream speech recognition. Systems condition neural separators or masking networks on enrollment speech, while other approaches use visual identity, microphone-array spatial cues, text, or neural signals; generative flows and autoregressive models offer alternatives to conventional separation. Current work increasingly addresses streaming and real conversations, where overlap, reverberation, noise, changing speakers, and differences between enrollment and target speech challenge models trained on simulated mixtures. A central tension is balancing intelligibility, speaker fidelity, perceptual quality, and latency, while ensuring objective metrics reflect real listening and recognition performance.