Abstract: A method for mitigating network abuse includes obtaining a first set of network traffic messages of network traffic currently received by a network service and determining, via a first model, whether network abuse is occurring based on the first set of network traffic messages. When the network abuse is occurring, the method includes obtaining a second set of current network traffic messages. The method also includes, for each network traffic message in the second set of network traffic messages, labeling, via a second model, the network traffic message as an abusing network traffic message or a non-abusing network traffic message. The method also includes generating, via a third model, at least one network traffic rule. Each network traffic rule, when implemented, reduces an effect of the abusing network traffic messages.

Abstract: The technology generally relates to spatial audio communication between devices. For example, a first device and a second device may be connected via a communication link. The first device may capture audio signals in an environment through two or more microphones. The first device may encode the captured audio with direction information. The first device may transmit the encoded audio via the communication link to the second device. The second device may decode the encoded audio to be output by one or more speakers of the second device. The second device may output the decoded audio to recreate positions of the captured audio signals.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for restarting a query using a token. One of the methods includes receiving, by a computer from a requesting device, a query; determining, using a data storage system, a current result responsive to the query; generating, using the current result, a restart token that represents operations performed to determine a plurality of results responsive to the query including the current result responsive to the query and that can be used to determine a new result responsive to the query that was not included in the plurality of results responsive to the query; and providing, to the requesting device, a message that includes a) first data for the restart token that represents operations performed to determine the plurality of results responsive to the query and b) second data for the current result responsive to the query.

Abstract: Techniques of controlling electronic devices using gestures use a wearable device on a user which translates, via a model, user movements into signals that both identify an electronic device to be controlled and a specific action to take with regard to that electronic device. The wearable device includes an inertial measurement unit (IMU) sensor and a photoplethysmography (PPG) sensor and measure six degrees of freedom (6DOF). The model is a convolutional neural network (CNN) that takes x, y, and z-acceleration signals generated by the IMU and PPG and places each acceleration component generated from each sensor in a separate channel. The CNN takes the input from each channel and generates a respective, separate model for each channel. The output at each of the stacked layers are combined in a fully connected layer to produce CNN output identifying an electronic device and a control for the electronic device.

Abstract: A battery pack includes a battery, a first temperature sensor configured to provide a first temperature value associated with a temperature of the battery, a heat source disposed proximate to the battery and configured to heat the battery, a second temperature sensor configured to provide a second temperature value associated with a temperature of the heat source, and a control board coupled to the first temperature sensor and the second temperature sensor, wherein the control board is configured to receive the first temperature value and the second temperature value. The control board is configured to compare the first temperature value and the second temperature value to determine a temperature gradient between the battery and the heat source and transmit an alert if the temperature gradient exceeds a first temperature gradient threshold.

Abstract: A method includes receiving a sequence of acoustic frames as input to an automatic speech recognition (ASR) model. The method also includes generating, by a first encoder, a first higher order feature representation for a corresponding acoustic frame. The method also includes generating, by a second encoder, a second higher order feature representation for a corresponding first higher order feature representation. The method also includes generating, by a language identification (ID) predictor, a language prediction representation based on a concatenation of the first higher order feature representation and the second higher order feature representation. The method also includes generating, by a first decoder, a first probability distribution over possible speech recognition hypotheses based on a concatenation of the second higher order feature representation and the language prediction representation.

Abstract: A method of detecting and responding to a visitor to a smart home environment via an electronic greeting system of the smart home environment, including determining that a visitor is approaching an entryway of the smart home environment; initiating a facial recognition operation while the visitor is approaching the entryway; initiating an observation window in response to the determination that a visitor is approaching the entryway; obtaining context information from one or more sensors of the smart home environment during the observation window; and at the end of the time window, initiating a response to the detected approach of the visitor based on the context information and/or an outcome of the facial recognition operation.

Abstract: Data are maintained in a distributed computing system that describe a graph. The graph represents relationships among items. The graph has a plurality of vertices that represent the items and a plurality of edges connecting the plurality of vertices. At least one vertex of the plurality of vertices includes a set of label values indicating the at least one vertex's strength of association with a label from a set of labels. The set of labels describe possible characteristics of an item represented by the at least one vertex. At least one edge of the plurality of edges includes a set of label weights for influencing label values that traverse the at least one edge. A label propagation algorithm is executed for a plurality of the vertices in the graph in parallel for a series of synchronized iterations to propagate labels through the graph.

Abstract: Methods, systems, apparatus, including computer programs encoded on computer storage media, for reclaiming storage space in a storage environment. In one aspect, the method includes actions of aggregating data that is indicative of access to one or more data objects, determining a future storage cost associated with each of a plurality of data objects, determining an access window for each of the plurality of data objects, identifying a data object based on (i) the future storage cost that satisfies a predetermined threshold and (ii) a data object access window, providing a notification to a user device that requests feedback from a user indicating whether the data object can be deleted, and in response to receiving data that indicates that the data object can be deleted, generating an instruction to cause deletion of the data object upon the expiration of the access window.

Abstract: A method for separating target speech from background noise contained in an input audio signal includes receiving the input audio signal captured by a user device, wherein the input audio signal corresponds to target speech of multiple words spoken by a target user and containing background noise in the presence of the user device while the target user spoke the multiple words in the target speech. The method also includes receiving a sequence of time markers input by the target user in cadence with the target user speaking the multiple words in the target speech, and correlating the sequence of time markers with the input audio signal to generate enhanced audio features that separate the target speech from the background noise in the input audio signal. The method also includes processing, using a speech recognition model, the enhanced audio features to generate a transcription of the target speech.

Abstract: A multichannel neural frontend speech enhancement model for speech recognition includes a speech cleaner, a stack of self-attention blocks each having a multi-headed self attention mechanism, and a masking layer. The speech cleaner receives, as input, a multichannel noisy input signal and a multichannel contextual noise signal, and generates, as output, a single channel cleaned input signal. The stack of self-attention blocks receives, as input, at an initial block of the stack of self-attention blocks, a stacked input including the single channel cleaned input signal and a single channel noisy input signal, and generates, as output, from a final block of the stack of self-attention blocks, an un-masked output. The masking layer receives, as input, the single channel noisy input signal and the un-masked output, and generates, as output, enhanced input speech features corresponding to a target utterance.

Abstract: A method for training a model includes obtaining a plurality of training samples. Each respective training sample of the plurality of training samples includes a respective speech utterance and a respective textual utterance representing a transcription of the respective speech utterance. The method includes training, using quantization aware training with native integer operations, an automatic speech recognition (ASR) model on the plurality of training samples. The method also includes quantizing the trained ASR model to an integer target fixed-bit width. The quantized trained ASR model includes a plurality of weights. Each weight of the plurality of weights includes an integer with the target fixed-bit width. The method includes providing the quantized trained ASR model to a user device.

Abstract: A method for optimal time-to-event (TTE) modeling includes obtaining a forecast request requesting performance of a TTE forecast forecasting an amount of time an event will occur after a starting point in time. The method includes obtaining a cutoff value representing an amount of time after the starting point in time that the event has not occurred. The method also includes forecasting, using an uncertainty forecasting model, the amount of time the event will occur after the starting point in time and updating the forecasted amount of time based on the cutoff value. The method also includes returning the updated forecasted amount of time the event will occur after the starting point in time.

Abstract: A method for speech conversion includes obtaining a speech conversion model configured to convert input utterances of human speech directly into corresponding output utterances of synthesized speech. The method further includes receiving a speech conversion request including input audio data corresponding to an utterance spoken by a target speaker associated with atypical speech and a speaker identifier uniquely identifying the target speaker. The method includes activating, using the speaker identifier, a particular sub-model for biasing the speech conversion model to recognize a type of the atypical speech associated with the target speaker identified by the speaker identifier.

Abstract: A method includes receiving a set of training utterances each including a non-synthetic speech representation of a corresponding utterance, and for each training utterance, generating a corresponding synthetic speech representation by using a voice conversion model. The non-synthetic speech representation and the synthetic speech representation form a corresponding training utterance pair. At each of a plurality of output steps for each training utterance pair, the method also includes generating, for output by a speech recognition model, a first probability distribution over possible non-synthetic speech recognition hypotheses for the non-synthetic speech representation and a second probability distribution over possible synthetic speech recognition hypotheses for the synthetic speech representation.

Abstract: A method includes generating, using an audio encoder, a higher-order feature representation for each acoustic frame in a sequence of acoustic frames; generating, using a decoder, based on the higher-order feature representation, a plurality of speech recognition hypotheses, each hypotheses corresponding to a candidate transcription of an utterance and having an associated first likelihood score; generating, using an external language model, for each speech recognition hypothesis, a second likelihood score; determining, using a learnable fusion module, for each speech recognition hypothesis, a set of fusion weights based on the higher-order feature representation and the speech recognition hypothesis; and generating, using the learnable fusion module, for each speech recognition hypothesis, a third likelihood score based on the first likelihood score, the second likelihood score, and the set of fusion weights, the audio encoder and decoder trained using minimum additive error rate training in the presence of t

Abstract: A method for generating emotionally intelligent responses to information seeking questions includes receiving audio data corresponding to a query spoken by a user and captured by an assistant-enabled device associated with the user, and processing, using a speech recognition model, the audio data to determine a transcription of the query. The method also includes performing query interpretation on the transcription of the query to identify an emotional state of the user that spoke the query, and an action to perform. The method also includes obtaining a response preamble based on the emotional state of the user and performing the identified action to obtain information responsive to the query. The method further includes generating a response including the obtained response preamble followed by the information responsive to the query.

Abstract: A method for training a generalized automatic speech recognition model for joint acoustic echo cancellation, speech enhancement, and voice separation includes receiving a plurality of training utterances paired with corresponding training contextual signals. The training contextual signals include a training contextual noise signal including noise prior to the corresponding training utterance, a training reference audio signal, and a training speaker vector including voice characteristics of a target speaker that spoke the corresponding training utterance. The operations also include training, using a contextual signal dropout strategy, a contextual frontend processing model on the training utterances to learn how to predict enhanced speech features. Here, the contextual signal dropout strategy uses a predetermined probability to drop out each of the training contextual signals during training of the contextual frontend processing model.

Abstract: A method of text-only and semi-supervised training for deliberation includes receiving training data including unspoken textual utterances that are each not paired with any corresponding spoken utterance of non-synthetic speech, and training a deliberation model that includes a text encoder and a deliberation decoder on the unspoken textual utterances. The method also includes receiving, at the trained deliberation model, first-pass hypotheses and non-causal acoustic embeddings. The first-pass hypotheses is generated by a recurrent neural network-transducer (RNN-T) decoder for the non-causal acoustic embeddings encoded by a non-causal encoder. The method also includes encoding, using the text encoder, the first-pass hypotheses generated by the RNN-T decoder, and generating, using the deliberation decoder attending to both the first-pass hypotheses and the non-causal acoustic embeddings, second-pass hypotheses.

Abstract: A method for detecting freeze words includes receiving audio data that corresponds to an utterance spoken by a user and captured by a user device associated with the user. The method also includes processing, using a speech recognizer, the audio data to determine that the utterance includes a query for a digital assistant to perform an operation. The speech recognizer is configured to trigger endpointing of the utterance after a predetermined duration of non-speech in the audio data. Before the predetermined duration of non-speech, the method includes detecting a freeze word in the audio data. In response to detecting the freeze word in the audio data, the method also includes triggering a hard microphone closing event at the user device. The hard microphone closing event prevents the user device from capturing any audio subsequent to the freeze word.