Abstract: A monitoring device has microphones, an ADC; a digital radio; and a processor with firmware. The firmware includes code for digitizing audio from the microphones into time-domain audio, performing FFT to provide frequency-domain audio, running a first neural network on time domain and frequency-domain audio to extract features, executing a classifier on the features to identify candidate events, and using the digital radio to upload candidate events and features. A pressure sensor awakens the processor from a low-power state. In particular embodiments, the first neural network is an embedded Gated Recurrent Unit having weights trained to extract features of use in the classifier; and candidate events include normal inhalation and exhalation breathing sounds, crackles, wheezes, coughs, snoring, gasping, choking, and speech sounds and in some embodiments heart sounds. A method of monitoring breathing during sleep includes attaching the device to, or embedding the device within, a pillow.

Abstract: A monitoring device has microphones, an ADC; a digital radio; and a processor with firmware. The firmware includes code for digitizing audio from the microphones into time-domain audio, performing FFT to provide frequency-domain audio, running a first neural network on time domain and frequency-domain audio to extract features, executing a classifier on the features to identify candidate events, and using the digital radio to upload candidate events and features. A pressure sensor awakens the processor from a low-power state. In particular embodiments, the first neural network is an embedded Gated Recurrent Unit having weights trained to extract features of use in the classifier; and candidate events include normal inhalation and exhalation breathing sounds, crackles, wheezes, coughs, snoring, gasping, choking, and speech sounds and in some embodiments heart sounds. A method of monitoring breathing during sleep includes attaching the device to, or embedding the device within, a pillow.