We develop environment-aware data-driven speech enhancement algorithms that take advantage of the widespread availability of microphones and can be deployed in ad hoc microphone networks of low-end devices. Our research is based on the observation, which was already validated in experiments, that in real-life acoustic environments, the recorded microphone signals have a distinct “fingerprint” of low-dimension that significantly simplifies the intricate physical acoustic models. We call the collection of the fingerprints from an environment, the “acoustic manifold”. We develop efficient and practical ways to construct such fingerprints for designated acoustic environments, and based on which, a new line of modern spatial acoustic processors.
See publications for more details.