Abstract: A system for configuring user-defined recognition patterns at an edge device using a hybrid cloud-edge device approach has a pattern recognition integrated circuit implementing a machine learning pattern recognizer that generates an event recognition output in response to an input thereto based upon pre-trained machine learning weights stored in a memory of the pattern recognition integrated circuit. A remote pattern recognition training service is in communication with a secondary user device receptive to a training input of the user-defined recognition patterns, and returns a set of training weights corresponding to the training input. An application interface connects the pattern recognition integrated circuit to the secondary user device, with the set of training weights returned to the secondary user device being transferable to the machine learning pattern recognizer for storage in the memory of the pattern recognition integrated circuit.

Abstract: A configurable monitoring and actioning system has one or more programmable edge devices each including a machine learning pattern recognizer, a sensor providing sensor input data to the pattern recognizer, and a memory storing pre-trained machine learning weight values for the pattern recognizer. Event detections are generated based upon evaluations of the sensor input data from the sensor against the pre-trained machine learning weight values. An application installable on a user device is in communication with each of the one or more programmable edge devices and executes predetermined actions based upon the event detection evaluations from the machine learning pattern recognizer.

Abstract: Discriminating between direct and machine-generated human voices is disclosed. A directly-generated voice audio sample from a human utterance and a machine-generated voice audio sample outputted by a loudspeaker from a pre-recording of another human utterance are captured on a microphone. Discriminative features between the directly-generated voice audio sample and the machine-generated voice audio sample are extracted with a machine learning classifier. A response to a command in the captured directly-generated voice audio sample or the captured machine-generated voice audio sample may be selectively generated.

Abstract: A deep learning training and inference system for a primary machine learning system has an automated data collection tool receptive to incoming input data from a sensor data source, and embeds one or more sensor data classifications associated with the incoming input data. A data augmentation tool is receptive to the input data from the automated data collection tool and generates an augmented input data set resulting from one or more predefined operations applied to the input data. An adaptive training tool is receptive to the augmented input data set to improve performance, with a new set of weight values being generated for the primary machine learning system. An inference tool is in communication with the adaptive training tool to receive the new set of weight values for an inference model simulator emulating a native hardware environment of the primary machine learning system.

Abstract: An always-on local action controller has one or more sensors each receptive to an external input. The respective external inputs are translatable to corresponding signals. One or more always-on data analytic neural network subsystems are each connected to a respective one of the sensors and are receptive to the signals outputted therefrom. An event detection is raised by a given one of the always-on data analytical neural network subsystems in response to a pattern of signal data corresponding to an event. A decision combiner is connected to each of the one or more always-on data analytic neural network subsystems, which generates an action signal based upon an aggregate of the events.

Abstract: A multi-stage noise suppression system for reducing noise components in a noisy input signal has a first stage neural network that estimates a noise power spectrum for the noisy input signal. A first set of gain values corresponding to the noise power spectrum is generated by the first stage neural network. A second stage neural network estimates clean signal power spectrum values, which are derived from an application of a second set of gain values generated as a function of the clean signal power spectrum values and a first stage reduced noise signal power spectrum values.

Abstract: A device wake-up system has one or more sensors each receptive to an external input. The respective external inputs are translatable to corresponding signals. One or more feature extractors connected to a respective one of the one or more sensors are receptive to the signals outputted from the sensors, and the feature data is associated with the signals being generated by the corresponding one of the one or more feature extractors. One or more inference circuits are connected to a respective one of the one or more feature extractors, and inference decisions are generated from patterns of the feature data generated by a corresponding one of the one or more feature extractors. A decision combiner is connected to each of the one or more inference circuits, and a wake signal is be generated based upon an aggregate of the inference decisions provided by the one or more inference circuits.

Abstract: A voice-activated system edge device cooperating with a remote command processor has a state machine defined by a listening mode state and a conversation monitoring mode state. The state machine transitions from the listening mode state to the conversation monitoring mode state in response to a wake word detection. A command accompanying the wake word is transmitted to the remote command processor for execution thereon. The conversation monitoring mode state is maintained for a conversation monitoring window time duration to receive a connection word accompanied by another command transmitted to the remote command processor for further execution thereon.

Abstract: A system can include a circuit holistically customized for the detection of commands in an audio or video input signal to meet certain application-specific requirements. The circuit can have a neural network topology that is hardwired to perform detection based on application-specific detection criteria. The hardwired custom circuit can provide improved energy efficiencies compared to similar functionality carried out using software and generic hardware modules. The system can also include a sound change trigger module and perform non-voiced sound detection.

Abstract: Systems and methods presented herein generally include multi-wake phrase detection executed on a single device utilizing multiple voice assistants. Systems and methods presented herein can further include continuously running a Voice Activity Detection (VAD) process which detects presence of human speech. The multi-wake phrase detection can activate when the VAD process detects human speech. Once activated, the multi-wake phrase detection can determine which (if any) of the wake phrases of the multiple voice assistants might be in the detected speech. Operation of the multi-wake phrase detection can have a low miss-rate. In some examples, operation of the multi-wake phrase detection can be granular to accomplish the low miss-rates at low power with a tolerance for false positives on wake phrase detection.

Abstract: A neural network parameter tuner has an auxiliary neural network receptive to an input data stream with signal components and noise components associated with ambient conditions. An ambient classification value is periodically derived from the input data stream based upon the noise components detected therein. A primary neural network receptive to the input data stream classifies the input data stream based upon an assigned detection threshold corresponding to the ambient classification value.

Abstract: A multi-stage selectable neural network noise suppression system has a first stage noise pattern selection neural network receptive to an input signal. An automatic noise classification is generated based upon an evaluation of the input signal. A noise suppression weight table stores one or more sets of automatic noise suppression weight values corresponding to the generated automatic noise classifications. A second stage noise pattern suppression neural network then selectively applies a specific automatic targeted noise suppression based upon the automatic noise suppression weight values. This approach balances performance quality and power/memory usage that efficiently tailors noise detection and suppression.

Abstract: Systems and methods are presented for recognizing and responding to voice commands at a local system and selectively streaming audio to a network-based computing system to recognize voice commands when the user provides a specific voice command to stream to the network-based computing system and/or when the user provides a voice command that is not recognizable by the local system.

Abstract: Systems and methods are presented herein that include utilizing a specialized circuit architecture for the purpose of locating lost items through voice command recognition. The system can be customized with minimal, low power components to continuously monitor for a locating phrase and produce an audible alert when the locating phrase is detected. The system can be included in a stand-alone portable device that can be attached to electronic and non-electronic portable objects, or the system can be embedded with other circuitry into a portable electronic device. The system can be powered by battery such that it is practical for the device to function as a locating device, continuously monitoring for the locating phrase. Voice recognition can be performed completely within the system, providing the benefit of privacy.

Abstract: A speech encoder that analyzes and classifies each frame of speech as being periodic-like speech or non-periodic like speech where the speech encoder performs a different gain quantization process depending if the speech is periodic or not. If the speech is periodic, the improved speech encoder obtains the pitch gains from the unquantized weighted speech signal and performs a pre-vector quantization of the adaptive codebook gain GP for each subframe of the frame before subframe processing begins and a closed-loop delayed decision vector quantization of the fixed codebook gain GC. If the frame of speech is non-periodic, the speech encoder may use any known method of gain quantization.

Abstract: In accordance with one aspect of the invention, a selector supports the selection of a first encoding scheme or the second encoding scheme based upon the detection or absence of the triggering characteristic in the interval of the input speech signal. The first encoding scheme has a pitch pre-processing procedure for processing the input speech signal to form a revised speech signal biased toward an ideal voiced and stationary characteristic. The pre-processing procedure allows the encoder to fully capture the benefits of a bandwidth-efficient, long-term predictive procedure for a greater amount of speech components of an input speech signal than would otherwise be possible. In accordance with another aspect of the invention, the second encoding scheme entails a long-term prediction mode for encoding the pitch on a sub-frame by sub-frame basis.

Abstract: There is provided a method for use by a speech encoder to encode an input speech signal.

Abstract: There is provided a method for use by a speech encoder to encode an input speech signal.

Abstract: There is provided a method for use by a speech encoder to encode an input speech signal.

Abstract: There is provided a voice activity detection method for indicating an active voice mode and an inactive voice mode. The method comprises receiving a first portion of an input signal; determining that the first portion of the input signal includes an active voice signal; indicating the active voice mode in response to the determining that the first portion of the input signal includes the active voice signal; receiving a second portion of the input signal immediately following the first portion of the input signal; determining that the second portion of the input signal includes an inactive voice signal; extending the indicating the active voice mode for a period of time after determining that the second portion of the input signal includes the inactive voice signal, wherein the period of time varies based on one or more conditions; and indicating the inactive voice mode after expiration of the period of time.