Selective Attention System (SAS): Device-Addressed Speech Detection for Real-Time On-Device Voice AI

The SAS project targets a critical friction point in voice AI: the 'intentionality gap'—deciding if a user is talking to a device or a person in a noisy environment without burning power on full ASR transcription. While the sequential modeling approach (SDAR) is technically sound and addresses the limitations of local classification, the project has zero stars and minimal external traction (8 days old). Defensibility is low because this specific capability is a 'holy grail' for Apple (Siri), Google (Assistant), and Amazon (Alexa); these frontier labs have massive proprietary datasets of multi-speaker interaction history that open-source projects cannot match. The displacement risk is high because next-generation end-to-end audio models (like GPT-4o or Gemini Live) are increasingly capable of inferring social context and 'addressivity' natively within the model architecture, potentially making standalone pre-ASR routing layers obsolete. From an investment perspective, this is a valuable research contribution but lacks the 'data gravity' or ecosystem lock-in required for a high defensibility score.