Sbir phase i: synchrony loop networks: a groundbreaking neuromorphic approach for live, unsupervised, one-shot source separation and diarization in complex auditory scenes

The National Science Foundation awarded Cambrya, Inc., $303,131 on August 15, 2026, under the NSF Technology, Innovation, and Partnerships program (CFDA 47.084) to develop a neuromorphic algorithm for real-time, unsupervised source separation and speaker diarization in complex acoustic environments.

The project delivers a novel neural learning approach combining local Hebbian learning with global network coordination mechanisms, paired with a heterogeneous network architecture that includes specialized layers for acoustic feature extraction, short-term auditory scene memory, and long-term sound characterization. The Phase I effort focuses on expanding the network's speech characterization and tracking layers to handle multi-speaker scenarios and developing prototype user controls for filtering and following selected sounds and voices. The innovation addresses a core limitation in artificial intelligence for edge device applications: rapid, unsupervised sensory learning that mirrors biological systems' ability to learn sounds within seconds of encountering them without labeled training data.

Performance occurs in Austin, Texas, with a period of performance ending July 31, 2027. No sub-awards are planned.

Abstract

The broader impact/commercial potential of this small business innovation research (sbir) phase i project is to develop a new technology with smart filtering functionality that separates each individual sound and voice in the acoustic field, and gives individuals easy control over which sound sources to attend to. The key innovation brought by this project is a new kind of algorithm that learns sounds and voices on its own within seconds of encountering them. This small business innovation research (sbir) phase i project brings to bear a new artificial learning technique designed for live, unsupervised, zero-shot sound source identification and separation. The two core innovations of this approach are a novel neural learning rule that combines elements of local hebbian learning with additional features that promote global network coordination; and a heterogeneous network architecture with specialized layers for acoustic feature extraction, short-term auditory scene memory, and long-term sound characterization. These advancements enable rapid and complex real-time sensory learning, a core feature of biological systems which has been an intractable problem for artificial intelligence (AI) that hampers AI usefulness for edge device applications. The phase i project focuses on expanding and refining the network?s speech characterization and tracking layers to handle multi-speaker situations, as well as developing prototype user controls for selecting sounds and voices to filter or follow. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.