Abstract: An apparatus for applying dynamic quantization of a neural network is described herein. The apparatus includes a scaling unit and a quantizing unit. The scaling unit is to calculate an initial desired scale factors of a plurality of inputs, weights and a bias and apply the input scale factor to a summation node. Also, the scaling unit is to determine a scale factor for a multiplication node based on the desired scale factors of the inputs and select a scale factor for an activation function and an output node. The quantizing unit is to dynamically requantize the neural network by traversing a graph of the neural network.

Abstract: An apparatus for applying dynamic quantization of a neural network is described herein. The apparatus includes a scaling unit and a quantizing unit. The scaling unit is to calculate an initial desired scale factors of a plurality of inputs, weights and a bias and apply the input scale factor to a summation node. Also, the scaling unit is to determine a scale factor for a multiplication node based on the desired scale factors of the inputs and select a scale factor for an activation function and an output node. The quantizing unit is to dynamically requantize the neural network by traversing a graph of the neural network.

Abstract: An apparatus for applying dynamic quantization of a neural network is described herein. The apparatus includes a scaling unit and a quantizing unit. The scaling unit is to calculate an initial desired scale factors of a plurality of inputs, weights and a bias and apply the input scale factor to a summation node. Also, the scaling unit is to determine a scale factor for a multiplication node based on the desired scale factors of the inputs and select a scale factor for an activation function and an output node. The quantizing unit is to dynamically requantize the neural network by traversing a graph of the neural network.

Abstract: Systems, apparatus and methods are described including operations for a flexible neural network accelerator.

Abstract: Techniques are disclosed for reliably masking speech commands directed to one or more computing devices to prevent the speech commands from being rendered. In some embodiments, each of the one or more computing devices includes components configured to generate acoustic data from ambient sound waves, process the acoustic data to identify a speech command sequence, and mask the speech command sequence from being rendered. At least some of the systems and methods disclosed herein monitor inbound audio at a fine grain level of detail. Working at this level of granularity enables the system and methods described herein to detect potential speech commands early within the user's utterance thereof and to discriminate quickly between true speech commands and other user utterances. These early detection and discrimination features, in turn, enable some embodiments to manage potential communication disruptions (e.g., jitter and/or latency) by modifying rates of audio prior to rendering.

Abstract: Techniques are disclosed for reliably masking speech commands directed to one or more computing devices to prevent the speech commands from being rendered. In some embodiments, each of the one or more computing devices includes components configured to generate acoustic data from ambient sound waves, process the acoustic data to identify a speech command sequence, and mask the speech command sequence from being rendered. At least some of the systems and methods disclosed herein monitor inbound audio at a fine grain level of detail. Working at this level of granularity enables the system and methods described herein to detect potential speech commands early within the user's utterance thereof and to discriminate quickly between true speech commands and other user utterances. These early detection and discrimination features, in turn, enable some embodiments to manage potential communication disruptions (e.g., jitter and/or latency) by modifying rates of audio prior to rendering.

Abstract: An apparatus for applying dynamic quantization of a neural network is described herein. The apparatus includes a scaling unit and a quantizing unit. The scaling unit is to calculate an initial desired scale factors of a plurality of inputs, weights and a bias and apply the input scale factor to a summation node. Also, the scaling unit is to determine a scale factor for a multiplication node based on the desired scale factors of the inputs and select a scale factor for an activation function and an output node. The quantizing unit is to dynamically requantize the neural network by traversing a graph of the neural network.

Abstract: Systems, apparatus and methods are described including operations for a flexible neural network accelerator.

Abstract: Technologies are described herein that allow a user to wake up a computing device operating in a low-power state and for the user to be verified by speaking a single wake phrase. Wake phrase recognition is performed by a low-power engine. In some embodiments, the low-power engine may also perform speaker verification. In other embodiments, the mobile device wakes up after a wake phrase is recognized and a component other than the low-power engine performs speaker verification on a portion of the audio input comprising the wake phrase. More than one wake phrases may be associated with a particular user, and separate users may be associated with different wake phrases. Different wake phrases may cause the device transition from a low-power state to various active states.

Abstract: Systems, apparatus and methods are described including operations for memory access via direct memory access engines, of a Gaussian Mixture Model Accelerator, corresponding to individual data streams.

Abstract: A disclosed speech processor includes a front end to receive a speech input and generate a feature vector indicative of a portion of the speech input and a Gaussian mixture (GMM) circuit to receive the feature vector, model any one of a plurality of GMM speech recognition algorithms, and generate a GMM score for the feature vector based on the GMM speech recognition algorithm modeled. In at least one embodiment, the GMM circuit includes a common compute block to generate feature a vector sum indicative of a weighted sum of differences squares between the feature vector and a mixture component of the GMM speech recognition algorithm. In at least one embodiment, the GMM speech recognition algorithm being modeled includes a plurality of Gaussian mixture components and the common compute block is operable to generate feature vector scores corresponding to each of the plurality of mixture components.

Abstract: Methods of enabling voice processing with minimal power consumption includes recording time-domain audio signal at a first clock frequency and a first voltage, and performing Fast Fourier Transform (FFT) operations on the time-domain audio signal at a second clock frequency to generate frequency-domain audio signal. The frequency domain audio signal may be enhanced to obtain better signal to noise ratio, through one or multiple filtering and enhancing techniques. The enhanced audio signal may be used to generate the total signal energy and estimate the background noise energy. Decision logic may determine from the signal energy and the background noise, the presence or absence of the human voice. The first clock frequency may be different from the second clock frequency.

Abstract: Technologies are described herein that allow a user to wake up a computing device operating in a low-power state and for the user to be verified by speaking a single wake phrase. Wake phrase recognition is performed by a low-power engine. in some embodiments, the low-power engine may also perform speaker verification. In other embodiments, the mobile device wakes up after a wake phrase is recognized and a component other than the low-power engine performs speaker verification on a portion of the audio input comprising the wake phrase, More than one wake phrases may be associated with a particular user, and separate users may be associated with different wake phrases. Different wake phrases may cause the device transition from a low-power state to various active states.

Abstract: Systems, apparatus and methods are described including operations for memory access via direct memory access engines, of a Gaussian Mixture Model Accelerator, corresponding to individual data streams.

Abstract: In some embodiments an accelerometer is mechanically coupled to a first device. The accelerometer obtains a vibration profile in response to a relative movement of the first device and a second device. A radio transmits an encrypted version of the vibration profile to the second device and receives an encrypted version of a vibration profile from the second device. A processor sets up a secure channel between the radio and the second device in which to exchange keys with the second device in order to decrypt the received encrypted vibration profile. The processor also decrypts the received encrypted vibration profile in response to at least one of the exchanged keys, compares the transmitted vibration profile with the received vibration profile and allows a sharing of resources with the second device if a match occurs between the transmitted vibration profile and the received vibration profile. Other embodiments are described and claimed.

Abstract: Technologies are described herein that allow a user to wake up a computing device operating in a low-power state and for the user to be verified by speaking a single wake phrase. Wake phrase recognition is performed by a low-power engine. In some embodiments, the low-power engine may also perform speaker verification. In other embodiments, the mobile device wakes up after a wake phrase is recognized and a component other than the low-power engine performs speaker verification on a portion of the audio input comprising the wake phrase. More than one wake phrases may be associated with a particular user, and separate users may be associated with different wake phrases. Different wake phrases may cause the device transition from a low-power state to various active states.

Abstract: In some embodiments a phase between a periodic spreading signal and an effective spreading signal modulating an interfering harmonic is determined, an amplitude of the interfering harmonic is estimated, and the interfering harmonic is canceled from a received signal. Other embodiments are described and claimed.

Abstract: A system, apparatus and method for efficient, low power, finite state transducer decoding. For example, one embodiment of a system for performing speech recognition comprises: a processor to perform feature extraction on a plurality of digitally sampled speech frames and to responsively generate a feature vector; an acoustic model likelihood scoring unit communicatively coupled to the processor over a communication interconnect to compare the feature vector against a library of models of various known speech sounds and responsively generate a plurality of scores representing similarities between the feature vector and the models; and a weighted finite state transducer (WFST) decoder communicatively coupled to the processor and the acoustic model likelihood scoring unit over the communication interconnect to perform speech decoding by traversing a WFST graph using the plurality of scores provided by the acoustic model likelihood scoring unit.

Abstract: Systems and methods may provide for using an audio front end of a mobile device to sampled audio from an audio signal during a first portion of a periodic detection window, and reducing a power consumption of one or more components of the audio front end during a second portion of the periodic detection window. Additionally, a determination may be made as to whether voice activity is present in the audio signal based at least in part on the sampled audio. In one example, the length of the first portion and the length of the second portion are defined by a duty cycle of the periodic detection window.

Abstract: Technologies are described herein that allow a user to wake up a computing device operating in a low-power state and for the user to be verified by speaking a single wake phrase. Wake phrase recognition is performed by a low-power engine. In some embodiments, the low-power engine may also perform speaker verification. In other embodiments, the mobile device wakes up after a wake phrase is recognized and a component other than the low-power engine performs speaker verification on a portion of the audio input comprising the wake phrase. More than one wake phrases may be associated with a particular user, and separate users may be associated with different wake phrases. Different wake phrases may cause the device transition from a low-power state to various active states.