Abstract: An electronic doorbell system can be configured to enable remote audio communications between a visitor at the doorbell and a user of a mobile computing device by exchanging speech-to-text and/or text-to-speech messages in real time. Audio captured by the visitor can be transcribed into text messages and sent to the user of the mobile device using a speech-to-text service. The user of the mobile device can send text messages to the doorbell for playback to the visitor by using a text-to-speech service. The system can also use artificial intelligence to detect the language spoken by the visitor for translating between a predetermined language of the user and the language of the visitor. The system can also include a camera for capturing video of the visitor for display to the mobile device simultaneous with exchanging text messages between the visitor and the user, such as during a live Session Initiation Protocol (SIP) communication.

Abstract: A method and an apparatus for determining a semantic matching degree, where the method includes acquiring a first sentence and a second sentence, dividing the first sentence and the second sentence into x and y sentence fragments, respectively, performing a convolution operation on word vectors in each sentence fragment of the first sentence and word vectors in each sentence fragment of the second sentence to obtain a three-dimensional tensor, performing integration or screening on adjacent vectors in the one-dimensional vectors of x rows and y columns, until the three-dimensional tensor is combined into a one-dimensional target vector, and determining a semantic matching degree between the first sentence and the second sentence according to the target vector.

Abstract: An artificial intelligence server for setting a language of a robot includes a communication unit and a processor. The communication unit is configured to communicate with the robot. The processor is configured to receive voice data for a control area from the robot, generate first language distribution information using the received voice data, receive event information for the control area, generate second language distribution information using the received event information, determine at least one major output language for the robot based on the generated first language distribution information and the generated second language distribution information, and transmit a control signal for setting the determined major output language to the robot.

Abstract: The present embodiments relate to analysing an image. An artificial deep neural net is pre-trained to classify images into a hierarchical system of multiple hierarchical classes. The pre-trained neural net is then adapted for one specific class, wherein the specific class is lower in the hierarchical system than an actual class of the image. The image is then processed by a forward pass through the adapted neural net to generate a processing result. An image processing algorithm is then used to analyse the processing result focused on features corresponding to the specific class.

Abstract: Systems and methods are described for object detection within a digital image using a hierarchical softmax function. The method may include applying a first softmax function of a softmax hierarchy on a digital image based on a first set of object classes that are children of a root node of a class hierarchy, then apply a second (and subsequent) softmax functions to the digital image based on a second (and subsequent) set of object classes, where the second (and subsequent) object classes are children nodes of an object class from the first (or parent) object classes. The methods may then include generating an object recognition output using a convolutional neural network (CNN) based at least in part on applying the first and second (and subsequent) softmax functions. In some cases, the hierarchical softmax function is the loss function for the CNN.

Abstract: A technique for training a neural network including an input layer, one or more hidden layers and an output layer, in which the trained neural network can be used to perform a task such as speech recognition. In the technique, a base of the neural network having at least a pre-trained hidden layer is prepared. A parameter set associated with one pre-trained hidden layer in the neural network is decomposed into a plurality of new parameter sets. The number of hidden layers in the neural network is increased by using the plurality of the new parameter sets. Pre-training for the neural network is performed.

Abstract: Discussed is an artificial intelligence (AI)-based voice sampling apparatus for providing a speech style, including a rhyme encoder configured to receive a user's voice, extract a voice sample, and analyze a vocal feature included in the voice sample, a text encoder configured to receive text for reflecting the vocal feature, a processor configured to classify the vocal feature of the voice sample input to the rhyme encoder according to a label, extract an embedding vector representing the vocal feature from the label, and generate a speech style from the embedding vector and apply the generated speech style to the text, and a rhyme decoder configured to output synthesized voice data in which the speech style is applied to the text by the processor.

Abstract: The present disclosure relates to systems and methods for speech signal processing on a signal to transcribe speech. In one implementation, the system may include a memory storing instructions and a processor configured to execute the instructions. The instructions may include instructions to receive the signal, determine if at least a portion of data in the signal is missing, and when at least a portion of data is missing: process the signal using a hidden Markov model to generate an output; using the output, calculate a set of possible contents to fill a gap due to the missing data portion, with each possible content having an associated probability; based on the associated probabilities, select one of the set of possible contents; and using the selected possible content, update the signal.

Abstract: Disclosed herein are neural networks for generating target classifications for an object from a set of input sequences. Each input sequence includes a respective input at each of multiple time steps, and each input sequence corresponds to a different sensing subsystem of multiple sensing subsystems. For each time step in the multiple time steps and for each input sequence in the set of input sequences, a respective feature representation is generated for the input sequence by processing the respective input from the input sequence at the time step using a respective encoder recurrent neural network (RNN) subsystem for the sensing subsystem that corresponds to the input sequence. For each time step in at least a subset of the multiple time steps, the respective feature representations are processed using a classification neural network subsystem to select a respective target classification for the object at the time step.

Abstract: Provided are a neural network device and a method of operation thereof. The neural network device for speaker recognition may include: a memory configured to store one or more instructions; and a processor configured to generate a trained second neural network by training a first neural network, for separating a mixed voice signal into individual voice signals by executing the one or more instructions, generate a second neural network by adding at least one layer to the trained first neural network, and generate a trained second neural network by training the second neural network, for separating the mixed voice signal into the individual voice signals and for recognizing a speaker of each of the individual voice signals.

Abstract: This specification describes computer-implemented methods and systems. One method includes receiving, by a neural network of a speech recognition system, first data representing a first raw audio signal and second data representing a second raw audio signal. The first raw audio signal and the second raw audio signal describe audio occurring at a same period of time. The method further includes generating, by a spatial filtering layer of the neural network, a spatial filtered output using the first data and the second data, and generating, by a spectral filtering layer of the neural network, a spectral filtered output using the spatial filtered output. Generating the spectral filtered output comprises processing frequency-domain data representing the spatial filtered output. The method still further includes processing, by one or more additional layers of the neural network, the spectral filtered output to predict sub-word units encoded in both the first raw audio signal and the second raw audio signal.

Abstract: Aspects of the subject disclosure may include, for example, embodiments provisioning a neural network comprising a plurality of layers. Further embodiments include provisioning a plurality of Markov logic state machines among the plurality of layers of the neural network resulting in a machine learning application. Additional embodiments include training the machine learning application using historical network video traffic resulting in a trained machine learning application. Also, embodiments include receiving current network video traffic. Embodiments include provisioning network resources to route the current network video traffic according to the trained machine learning application. Other embodiments are disclosed.

Abstract: A method includes: capturing one or more images of an unorganized collection of items inside a first machine; determining one or more item types of the unorganized collection of items from the one or more images, comprising: dividing a respective image in the one or more images into a respective plurality of sub-regions; performing feature detection on the respective plurality of sub-regions to obtain a respective plurality of regional feature vectors, wherein a regional feature vector for a sub-region indicates characteristics for a plurality of predefined local item features for the sub-region; generating an integrated feature vector by combining the respective plurality of regional feature vectors; and applying a plurality of binary classifiers to the integrated feature vector; and selecting a machine setting for the first machine based on the determined one or more clothes type in the unorganized collection of items.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for personalized speaker verification are provided. One of the methods includes: obtaining first speech data of a speaker as a positive sample and second speech data of an entity different from the speaker as a negative sample; feeding the positive sample and the negative sample to a first model for determining voice characteristics to correspondingly output a positive voice characteristic and a negative voice characteristic of the speaker; obtaining a gradient based at least on the positive voice characteristic and the negative voice characteristic; and feeding the gradient to the first model to update one or more parameters of the first model to obtain a second model for personalized speaker verification.

Abstract: Approaches for the automatic segmentation of magnetic resonance (MR) images. Machine learning models segment images to identify image features in consecutive frames at different levels of resolution. A neural network block is applied to groups of MR images to produce primary feature maps at two or more levels of resolution. The images in a given group of MR images may correspond to a cycle and have a temporal order. A second RNN block is applied to the primary feature maps to produce two or more output tensors at corresponding levels of resolution. A segmentation block is applied to the two or more output tensors to produce a probability map for the MR images. The first neural network block may be a convolutional neural network (CNN) block. The second neural network block may be a convolutional long short-term (LSTM) block.

Abstract: Provided are an electronic device and a control method. The electronic device comprises: a storage unit for storing a user-based dictionary; an input unit for receiving an input sentence including a user-specific word and at least one word learned by a neural network-based language model; and a processor for determining a concept category of the user-specific word on the basis of semantic information of the input sentence, adding the user-specific word to the user-based dictionary to perform update, and when text corresponding to semantic information of the at least one learned word is input, providing the user-specific word as an autocomplete recommendation word which can be input subsequent to the text.

Abstract: Techniques for avoiding wake word self-triggering are provided. In one embodiment, an electronic device can receive an audio-out signal to be output as audio via a speaker of the device and can attempt to recognize a wake word in the audio-out signal using a first recognizer. If the wake word is recognized in the audio-out signal, the electronic device can further determine whether a wake word match is made using a second recognizer with respect to a mic-in audio signal captured via a microphone of the device at approximately the same time that the audio-out signal is output via the speaker. If so, the electronic device can ignore the wake word match made using the second recognizer.

Abstract: Embodiments of the present systems and methods may provide techniques for extracting vocal features from voice signals to determine an emotional or mental state of one or more persons, such as to determine a risk of suicide and other mental health issues. For example, as a person's mental state may indirectly alters his or her speech, suicidal risk in, for example, hotline calls, may be determined through speech analysis. In embodiments, such techniques may include preprocessing of the original recording, vocal feature extraction, and prediction processing. For example, in an embodiment, a computer-implemented method of determining an emotional or mental state of a person, the method comprising acquiring an audio signal relating to a conversation including the person, extracting signal components relating to an emotional or mental state of at least the person, and outputting information characterizing the extracted emotional or mental state of the person.

Abstract: A method and system of creating a model for large scale data analytics is provided. Training data is received in a form of a data matrix X and partitioned into a plurality of partitions. A random matrix T is generated. A feature matrix is determined based on multiplying the partitioned training data by the random matrix T. A predicted data {tilde over (y)} is determined for each partition via a stochastic average gradient (SAG) of each partition. A number of SAG values is reduced based on a number of rows n in the data matrix X. For each iteration, a sum of the reduced SAG values is determined, as well as a full gradient based on the sum of the reduced SAG values from all rows n, by distributed parallel processing. The model parameters w are updated based on the full gradient for each partition.

Abstract: Disclosed are a speech recognition method, a device and a computer readable storage medium. The speech recognition method comprises: performing acoustic characteristic extraction on an input voice, so as to obtain an acoustic characteristic (S11); acquiring an acoustic model, a parameter of the acoustic model being a binarization parameter (S12); and according to the acoustic characteristic and the acoustic model, performing speech recognition (S13).

Abstract: Provided is a voice conversation method using a bi-directional LSTM (Bi-LSTM) memory network.

Abstract: Technologies for optimization of machine learning training include a computing device to train a machine learning network with a training algorithm that is configured with configuration parameters. The computing device may perform many training instances in parallel. The computing device captures a time series of partial accuracy values from the training. Each partial accuracy value is indicative of machine learning network accuracy at an associated training iteration. The computing device inputs the configuration parameters to a feed-forward neural network to generate a representation and inputs the representation to a recurrent neural network. The computing device trains the feed-forward neural network and the recurrent neural network against the partial accuracy values. The computing device optimizes the feed-forward neural network and the recurrent neural network to determine optimized configuration parameters.

Abstract: An audio processing apparatus, comprising: a first receiver configured to receive one or more audio signals derived from one or more microphones, the one or more audio signals comprising a speech component received from a user and a first noise component transmitted by a first device; a second receiver configured to receive over a network and from the first device, first audio data corresponding to the first noise component; one or more processors configured to: remove the first noise component from the one or more audio signals using the first audio data to generate a first processed audio signal; and perform speech recognition on the first processed audio signal to generate a first speech result.

Abstract: Systems, methods, and devices for speech transformation and generating synthetic speech using deep generative models are disclosed. A method of the disclosure includes receiving input audio data comprising a plurality of iterations of a speech utterance from a plurality of speakers. The method includes generating an input spectrogram based on the input audio data and transmitting the input spectrogram to a neural network configured to generate an output spectrogram. The method includes receiving the output spectrogram from the neural network and, based on the output spectrogram, generating synthetic audio data comprising the speech utterance.

Abstract: Techniques recognize reducible contents in data to be written. The techniques involve receiving information related to data to be written, the information indicating that the data to be written comprises reducible contents, the reducible contents comprising data with a first reduction pattern. The techniques further involve recognizing the reducible contents in the data to be written based on the information. The techniques further involve reducing the reducible contents based on the first reduction pattern. With such techniques, active I/O pattern recognition with communication between applications and storage devices may be accomplished. In addition, with such techniques, it is easy/simple to expand recognizable new patterns, and I/O pattern limitations in standard approaches no longer exist.

Abstract: A method for training an endpointer model includes short-form speech utterances and long-form speech utterances. The method also includes providing a short-form speech utterance as input to a shared neural network, the shared neural network configured to learn shared hidden representations suitable for both voice activity detection (VAD) and end-of-query (EOQ) detection. The method also includes generating, using a VAD classifier, a sequence of predicted VAD labels and determining a VAD loss by comparing the sequence of predicted VAD labels to a corresponding sequence of reference VAD labels. The method also includes, generating, using an EOQ classifier, a sequence of predicted EOQ labels and determining an EOQ loss by comparing the sequence of predicted EOQ labels to a corresponding sequence of reference EOQ labels. The method also includes training, using a cross-entropy criterion, the endpointer model based on the VAD loss and the EOQ loss.

Abstract: According to some aspects, a method of searching for content in response to a user voice query is provided. The method may comprise receiving the user voice query, performing speech recognition to generate N best speech recognition results comprising a first speech recognition result, performing a supervised search of at least one content repository to identify one or more supervised search results using one or more classifiers that classify the first speech recognition result into at least one class that identifies previously classified content in the at least one content repository, performing an unsupervised search of the at least one content repository to identify one or more unsupervised search results, wherein performing the unsupervised search comprises performing a word search of the at least one content repository, and generating combined results from among the one or more supervised search results and the one or more unsupervised search results.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining, by a first sequence-training speech model, a first batch of training frames that represent speech features of first training utterances; obtaining, by the first sequence-training speech model, one or more first neural network parameters; determining, by the first sequence-training speech model, one or more optimized first neural network parameters based on (i) the first batch of training frames and (ii) the one or more first neural network parameters; obtaining, by a second sequence-training speech model, a second batch of training frames that represent speech features of second training utterances; obtaining one or more second neural network parameters; and determining, by the second sequence-training speech model, one or more optimized second neural network parameters based on (i) the second batch of training frames and (ii) the one or more second neural network parameters.

Abstract: A speech recognition device includes: an acoustic model based on an End-to-End neural network responsive to an observed sequence formed of prescribed acoustic features obtained from a speech signal by feature extracting unit, for calculating probability of the observed sequence being a certain symbol sequence; and a decoder responsive to a symbol sequence candidate, for decoding a speech signal by a WFST based on a posterior probability of each of word sequences corresponding to the symbol sequence candidate, probabilities calculated by the acoustic model for symbol sequences selected based on an observed sequence, and a posterior probability of each of the plurality of symbol sequences.

Abstract: The present teaching relates to method, system, and medium for generating training data for generating a relationship identification model. Sentences are received as input. Each of the sentences is aligned with a fact previously stored to create an alignment. Confidence scores for the alignments are computed and then used, together with the alignments to train a relationship identification model.

Abstract: Techniques for adapting a trained neural network acoustic model, comprising using at least one computer hardware processor to perform: generating initial speaker information values for a speaker; generating first speech content values from first speech data corresponding to a first utterance spoken by the speaker; processing the first speech content values and the initial speaker information values using the trained neural network acoustic model; recognizing, using automatic speech recognition, the first utterance based, at least in part on results of the processing; generating updated speaker information values using the first speech data and at least one of the initial speaker information values and/or information used to generate the initial speaker information values; and recognizing, based at least in part on the updated speaker information values, a second utterance spoken by the speaker.

Abstract: A method for temporal logic fusion can include steps of: receiving a plurality of inputs for a plurality of behavior classes, the inputs consisting of single- or multi-dimensional states sampled over time; computing a distance metric pairwise among the inputs, the computing being performed using dynamic time warping; mapping the high-dimension input signals into a 2-dimensional (2-D) space using t-distributed Stochastic Neighbor Embedding, the pairwise computation from the computing step being used as the distance metric required to perform this mapping; clustering the high-dimension input signals in the 2-D space via a k-means clustering algorithm; generating a signal temporal logic (STL) expression that distinguishes between a cluster in a behavior class and all high-dimension input signals not in that behavior class; and repeating the generating step for each cluster in that behavior class.

Abstract: A method includes that for each model from multiple models, evaluating a model prediction accuracy based on a dataset of a user over a first time duration. The dataset includes a sequence of actions with corresponding contexts based on electronic device interactions. Each model is trained to predict a next action at a time point within the first time duration, based on a first behavior sequence over a first time period from the dataset before the time point, a second behavior sequence over a second time period from the dataset before the time point, and context at the time point. A model is selected from the multiple models based on its model prediction accuracy for the user based on a domain. An action to be initiated at a later time using an electronic device of the user is recommended using the selected model during a second time duration.

Abstract: A system and computer-executable program code for accelerated rescoring with recurrent neural net language models (RNNLMs) on hybrid CPU/GPU computer systems utilizes delayed, frame-wise (or layer-wise) batch dispatch of RNNLM scoring tasks to the GPU(s), while performing substantially all other tasks on the CPU(s).

Abstract: Technologies for detecting minority events are disclosed. By performing a guided hierarchical classification algorithm with a decision tree structure and grouping the minority class(es) in with some of the majority classes, large majority classes may be separated from a minority class without requiring good detection of the minority events by themselves. The decision tree structure may be used only for the purpose of identifying if the data sample in question is a member of a minority class. If it is determined that it is not, a primary classification algorithm may be used. With this approach, the guided hierarchical classification algorithm need not perform as well as the primary classification algorithm for the majority events, but may provide improved detection for minority events.

Abstract: Computer-implemented techniques can include obtaining, by a client computing device, a digital media item and a request for a processing task on the digital item and determining a set of operating parameters based on (i) available computing resources at the client computing device and (ii) a condition of a network. Based on the set of operating parameters, the client computing device or a server computing device can select one of a plurality of artificial neural networks (ANNs), each ANN defining which portions of the processing task are to be performed by the client and server computing devices. The client and server computing devices can coordinate processing of the processing task according to the selected ANN. The client computing device can also obtain final processing results corresponding to a final evaluation of the processing task and generate an output based on the final processing results.

Abstract: A knowledge management system includes: a default knowledge system including: a knowledge system and a knowledge database in communication with the knowledge system; and a knowledge store in communication with the default knowledge system and including: a taxonomy amendment, an annotation amendment, a canonicalization amendment, an ecosystem amendment, a term amendment, and a phrase amendment.

Abstract: A method of attention-based end-to-end (E2E) automatic speech recognition (ASR) training, includes performing cross-entropy training of a model, based on one or more input features of a speech signal, performing beam searching of the model of which the cross-entropy training is performed, to generate an n-best hypotheses list of output hypotheses, and determining a one-best hypothesis among the generated n-best hypotheses list. The method further includes determining a character-based gradient and a word-based gradient, based on the model of which the cross-entropy training is performed and a loss function in which a distance between a reference sequence and the determined one-best hypothesis is maximized, and performing backpropagation of the determined character-based gradient and the determined word-based gradient to the model, to update the model.

Abstract: A method (and structure and computer product) for learning Out-of-Vocabulary (OOV) words in an Automatic Speech Recognition (ASR) system includes using an Acoustic Word Embedding Recurrent Neural Network (AWE RNN) to receive a character sequence for a new OOV word for the ASR system, the RNN providing an Acoustic Word Embedding (AWE) vector as an output thereof. The AWE vector output from the AWE RNN is provided as an input into an Acoustic Word Embedding-to-Acoustic-to-Word Neural Network (AWE?A2W NN) trained to provide an OOV word weight value from the AWE vector. The OOV word weight is inserted into a listing of Acoustic-to-Word (A2W) word embeddings used by the ASR system to output recognized words from an input of speech acoustic features, wherein the OOV word weight is inserted into the A2W word embeddings list relative to existing weights in the A2W word embeddings list.

Abstract: Exemplary embodiments relate to improvements to neural networks for translation and other sequence-to-sequence tasks. A convolutional neural network may include multiple blocks, each having a convolution layer and gated linear units; gating may determine what information passes through to the next block level. Residual connections, which add the input of a block back to its output, may be applied around each block. Further, an attention may be applied to determine which word is most relevant to translate next. By applying repeated passes of the attention to multiple layers of the decoder, the decoder is able to work on the entire structure of a sentence at once (with no temporal dependency). In addition to better accuracy, this configuration is better at capturing long-range dependencies, better models the hierarchical syntax structure of a sentence, and is highly parallelizable and thus faster to run on hardware.

Abstract: According to an embodiment, a training device trains a neural network that outputs a posterior probability that an input signal belongs to a particular class. An output layer of the neural network includes N units respectively corresponding to classes and one additional unit. The training device includes a propagator, a probability calculator, and an updater. The propagator supplies a sample signal to the neural network and acquires (N+1) input values for each unit at the output layer. The probability calculator supplies the input values to a function to generate a probability vector including (N+1) probability values respectively corresponding to the units at the output layer. The updater updates a parameter included in the neural network in such a manner to reduce an error between a teacher vector including (N+1) target values and the probability vector. A target value corresponding to the additional unit is a predetermined constant value.

Abstract: A method for transferring acoustic knowledge of a trained acoustic model (AM) to a neural network (NN) includes reading, into memory, the NN and the AM, the AM being trained with target domain data, and a set of training data including a set of phoneme data, the set of training data being data obtained from a domain different from a target domain for the target domain data, inputting training data from the set of training data into the AM, calculating one or more posterior probabilities of context-dependent states corresponding to phonemes in a phoneme class of a phoneme to which each frame in the training data belongs, and generating a posterior probability vector from the one or more posterior probabilities, as a soft label for the NN, and inputting the training data into the NN and updating the NN, using the soft label.

Abstract: Disclosed is a computer implemented method for malware detection that analyses a file on a per packet basis. The method receives a packet of one or more packets associated a file, and converting a binary content associated with the packet into a digital representation and tokenizing plain text content associated with the packet. The method extracts one or more n-gram features, an entropy feature, and a domain feature from the converted content of the packet and applies a trained machine learning model to the one or more features extracted from the packet. The output of the machine learning method is a probability of maliciousness associated with the received packet. If the probability of maliciousness is above a threshold value, the method determines that the file associated with the received packet is malicious.

Abstract: An information processing apparatus includes: a speech obtainer which obtains speech of a user; a first controller which, when the first controller recognizes that the speech obtained by the speech obtainer is a first activation word, outputs a speech signal corresponding to the first activation word; and a second controller. In the first speech transmission process in which the speech signal of the speech obtained by speech obtainer is transmitted to the VPA cloud server, the first controller determines whether to output a speech signal corresponding to a second activation word to the second controller based on a predetermined priority level when the first controller recognizes that the speech obtained by the speech obtainer indicates the second activation word for causing the second controller to start a second speech transmission process.

Abstract: An audio device with at least one microphone adapted to receive sound from a sound field and create an output, and a processing system that is responsive to the output of the microphone. The processing system is configured to use a signal processing algorithm to detect a wakeup word, and modify the signal processing algorithm that is used to detect the wakeup word if the sound field changes.

Abstract: In one embodiment, a computer program product for optimizing core utilization in a neurosynaptic network includes a computer readable storage medium having program instructions embodied therewith, where the computer readable storage medium is not a transitory signal per se, and where the program instructions are executable by a processor to cause the processor to perform a method including identifying, by the processor, one or more unused portions of a neurosynaptic network, and for each of the one or more unused portions of the neurosynaptic network, disconnecting, by the processor, the unused portion from the neurosynaptic network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a second model to approximate the output of a first model to classify, according to a classification scheme, image data received as input, and after the second model is trained accessing map data that specifies a plurality of geographic locations, and for each geographic location associated with an entity for each image of the one or more images that depict the entity located at the geographic location, providing the image to the second model to generate an embedding for the image, associating each of the one or more embeddings generated by the second model with the geographic location, and storing, in a database, location data specifying the geographic location, the associated one or more embeddings, and data specifying the entity, as an associated entity entry for the entity.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for keyword spotting. One of the methods includes training, by a keyword detection system, a convolutional neural network for keyword detection by providing a two-dimensional set of input values to the convolutional neural network, the input values including a first dimension in time and a second dimension in frequency, and performing convolutional multiplication on the two-dimensional set of input values for a filter using a frequency stride greater than one to generate a feature map.

Abstract: A speech wakeup method, apparatus, and electronic device are disclosed in embodiments of this specification. The method includes: inputting speech data to a speech wakeup model trained with general speech data; and outputting, by the speech wakeup model, a result for determining whether to execute speech wakeup, wherein the speech wakeup model includes a Deep Neural Network (DNN) and a Connectionist Temporal Classifier (CTC).

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a Q network used to select actions to be performed by an agent interacting with an environment. One of the methods includes obtaining a plurality of experience tuples and training the Q network on each of the experience tuples using the Q network and a target Q network that is identical to the Q network but with the current values of the parameters of the target Q network being different from the current values of the parameters of the Q network.