CA-NeuroSpex: Context-Informed Autoregressive Neuro-Guided Speaker Extraction.

Area of Science:

• Neuroscience
• Speech Processing
• Machine Learning

Background:

• Neuro-guided target speaker extraction (TSE) uses brain signals to isolate speech.
• Existing methods neglect the crucial role of temporal context in speech perception.

Purpose of the Study:

• To develop a novel context-informed end-to-end TSE framework (CA-NeuroSpex).
• To integrate temporal context with neural signals for enhanced speech extraction.

Main Methods:

• Introduced a specialized speech-context encoder for integrating past speech information.
• Employed an autoregressive feedback mechanism to incorporate temporal context.
• Developed a gating mechanism for dynamically fusing neural and contextual cues.

Main Results:

• CA-NeuroSpex demonstrated improved target speaker extraction performance.
• The framework effectively integrates overlapping speech using contextual cues.
• Dynamic fusion of contextual and neural information proved robust.

Conclusions:

• Combining temporal context with neural guidance enhances target speaker extraction.
• CA-NeuroSpex offers a robust solution for separating attended speech in complex acoustic scenes.
• The proposed method advances the field of neuro-guided speech processing.