Microsoft Technology Licensing, LLC

Shuayb M Zarar

Ivan Jelev Tashev

Jun Zhang

Amol Ashok Ambardekar

Matthai Philipose

Amit Das

Yan-Hui Tu

Jong Hwan Ko

Chin-Hui Lee

Han Zhao

Systems, methods, and computer-readable storage devices are disclosed for improved real-time audio processing. One method including: constructing a deep neural network model, including a plurality of at least one-bit neurons, configured to output a predicted label of audio data, the plurality of at least one-bit neurons arranged in a plurality of layers, including at least one hidden layer, and being connected by a plurality of connections, each connection having at least a one-bit weight, wherein one or both of the plurality of at least one-bit neurons and the plurality of connections have a reduced bit precision; receiving a training data set, the training data set including audio data; training the deep neural network model using the training data set; and outputting a trained deep neural network model configured to output a predicted label of real-time audio data.

Systems, methods, and computer-readable media for improved real-time audio processing

A method of performing matrix computations includes receiving a compression-encoded matrix including a plurality of rows. Each row of the compression-encoded matrix has a plurality of defined element values and, for each such defined element value, a schedule tag indicating a schedule for using the defined element value in a scheduled matrix computation. The method further includes loading the plurality of rows of the compression-encoded matrix into a corresponding plurality of work memory banks, and providing decoded input data to a matrix computation module configured for performing the scheduled matrix computation. For each work memory bank, a next defined element value and a corresponding schedule tag are read. If the schedule tag meets a scheduling condition, the next defined element value is provided to the matrix computation module. Otherwise, a default element value is provided to the matrix computation module.

Compression-encoding scheduled inputs for matrix computations

A computer system comprising a scheduled computation module, a work memory storage device, and a controller. The scheduled computation module is configured to receive and process data values according to a predetermined access pattern. The work memory storage device includes one or more work memory banks. The controller is configured to, based on scheduling information associated with the predetermined access pattern, (1) provide data values held by the one or more work memory banks to the scheduled computation module, and (2) selectively control a power state of the one or more work memory banks.

Selectively controlling memory power for scheduled computations

A facility for processing output from a network of mechanical sensors is described. The facility accesses time-series data outputted by the network of sensors. The facility applies to the accessed time-series data a trained autoencoder to obtain a version of the accessed time-series data in which noise present in the accessed time-series data is at least partially suppressed. The facility stores the obtained version of the accessed time-series data, such as in order to perform human activity recognition against the obtained version of the accessed time-series data.

Learning-based noise reduction in data produced by a network of sensors, such as one incorporated into loose-fitting clothing worn by a person

Systems, methods, and computer-readable storage devices are disclosed for improved real-time audio processing. One method including: receiving audio data including a plurality of frames having a plurality of frequency bins; calculating, for each frequency bin, an approximate speech signal estimation based on the plurality of frames; calculating, for each approximate speech signal estimation, a clean speech estimation and at least one additional target including an ideal ratio mask using a trained neural network model; and calculating, for each frequency bin, a final clean speech estimation using the calculated at least one additional target including the calculated ideal ratio mask and the calculated clean speech estimation.

An ultrasonic gesture recognition system is provided that recognizes gestures based on analysis of return signals of an ultrasonic pulse that is reflected from a gesture. The system transmits an ultrasonic chirp and samples a microphone array at sample intervals to collect a return signal for each microphone. The system then applies a beamforming technique to frequency domain representations of the return signals to generate an acoustic image with a beamformed return signal for multiple directions. The system then generates a feature image from the acoustic images to identify, for example, distance or depth from the microphone array to the gesture for each direction. The system then submits the feature image to a deep learning system to classify the gesture.

Ultrasonic based gesture recognition

A system is provided that employs a statistical approach to semi-supervised speech enhancement with a low-order non-negative matrix factorization ('NMF'). The system enhances noisy speech based on multiple dictionaries with dictionary atoms derived from the same clean speech samples and generates an enhanced speech representation of the noisy speech by combining, for each dictionary, a clean speech representation of the noisy speech generated based on a NMF using the dictionary atoms of the dictionary. The system generates frequency-domain ('FD') clean speech sample representations of the clean speech samples, for example, using a Fourier transform. To generate each dictionary, the system generates a dictionary-unique initialization of the dictionary atoms and the activations and performs a NMF of the FD clean speech samples.

Growing community of inventors

Redmond, WA, United States of America

Shuayb M Zarar