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 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 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 computer-implemented method includes receiving a sequence of acoustic frames corresponding to an utterance and generating a reference speaker embedding for the utterance. The method also includes receiving a target speaker embedding for a target speaker and generating feature-wise linear modulation (FiLM) parameters including a scaling vector and a shifting vector based on the target speaker embedding. The method also includes generating an affine transformation output that scales and shifts the reference speaker embedding based on the FiLM parameters. The method also includes generating a classification output indicating whether the utterance was spoken by the target speaker based on the affine transformation output.

Abstract: A method for time series forecasting includes receiving a time series forecasting query from a user requesting the data processing hardware to perform a plurality of time series forecasts. Each time series forecast is a forecast of future data based on respective current data. Simultaneously, for each time series forecast of the plurality of time series forecasts requested by the time series forecasting query, the method includes training a plurality of models for the respective time series forecast. The method also includes determining which model of the plurality of models best fits the respective time series forecast and forecasting the future data based on the determined best fitting model and the respective current data. The method also includes returning, to the user, the forecasted future data for each of the plurality of time series forecasts request by the timer series forecasting query.

Abstract: A system comprising an encoder neural network, a scene structure decoder neural network, and a motion decoder neural network. The encoder neural network is configured to: receive a first image and a second image; and process the first image and the second image to generate an encoded representation of the first image and the second image. The scene structure decoder neural network is configured to process the encoded representation to generate a structure output characterizing a structure of a scene depicted in the first image. The motion decoder neural network configured to process the encoded representation to generate a motion output characterizing motion between the first image and the second image.

Abstract: A computer-implemented method includes receiving audio data corresponding to a spoken statement by a user and converting the audio data into a meaning representation of the natural language description for creating the application. The spoken statement includes a natural language description for creating an application. Moreover, the meaning representation includes one or more inferences made from the natural language description for creating the application. The method further includes publishing a functional version of the application based on the natural language description. The functional version of the application is executable to perform operations specified by the natural language description of the application.

Abstract: Methods, systems, and media for detecting a presentation of media content on a display device are provided. In accordance with some implementations, methods for detecting a presentation of media content on a display device are provided, the methods comprising: detecting, using a light sensor, light levels in the light sensor's surroundings; generating a signal representing the light levels; detecting, using a hardware processor, at least one variation in light levels indicative of a presentation of a video scene based on the signal; detecting at least one variation in light levels indicative of a scene change subsequent to the video scene based on the signal; and determining that media content is being presented on a display device in response to detecting the variation in light levels indicative of the presentation of the video scene and the variation in light levels indicative of the scene change.

Abstract: Methods and systems, including computer programs encoded on a computer storage medium. In one aspect, a method includes the actions of receiving a request to perform convolutional computations for a neural network on a hardware circuit having a matrix computation unit, the request specifying the convolutional computation to be performed on a feature tensor and a filter and padding applied to the feature tensor prior to performing the convolutional computation; and generating instructions that when executed by the hardware circuit cause the hardware circuit to perform operations comprising: transferring feature tensor data from a main memory of the hardware circuit to a scratchpad memory of the hardware circuit; and repeatedly performing the following operations: identifying a current subset of the feature tensor; and determining whether a memory view into the scratchpad memory for the current subset is consistent with a memory view of the current subset in the main memory.

Abstract: Systems and methods of routing are provided. In the system, one or more processors determine that a packet is to be transmitted to a destination. In one or more aspects of the system, the one or more processors select a next port to be used for transmitting the packet by selecting a set of ports among a plurality of ports based on a static weight configuration associated with each port. The next port may be selected from the set of ports based on a number of hops required to reach the destination from each port and based on an estimated latency from each port to the destination. The one or more processors may then route the packet through the selected next port.

Abstract: A system for simulating network configurations includes data processing hardware and memory hardware in communication with the data processing hardware. The memory hardware stores instructions that when executed on the data processing hardware cause the data processing hardware to perform operations. The operations includes receiving one or more parameter changes for a production network model of a network. The operations also include generating a simulation network model including the one or more parameter changes. Another operation includes analyzing the simulated network flow within the simulation network model. The operations also include generating a report including an impact of the parameter changes on the network. The operations may also include receiving a production network log including a recorded workflow for the production network model and simulating the production workflow of the production network log within the simulation network model to generate a simulated network log.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for scheduling operations on a machine-learning accelerator having multiple tiles. The apparatus includes a processor having a plurality of tiles and scheduling circuitry that is configured to select a respective input activation for each tile of the plurality of tiles from either an activation line for the tile or a delay register for the activation line.

Abstract: Systems and methods for performing inference for word or wordpiece tokenization are disclosed using a left-to-right longest-match-first greedy process. In some examples, the vocabulary may be organized into a trie structure in which each node includes a precomputed token or token ID and a fail link, so that the tokenizer can parse the trie in a single pass to generate a list of only those tokens or token IDs that correspond to the longest matching vocabulary entries in the sample string, without the need for backtracking. In some examples, the vocabulary may be organized into a trie in which each node has a fail link, and any node that would share token(s) or token_ID(s) of a preceding node is instead given a prev_match link that points back to a chain of nodes with those token(s) or token_ID(s).

Abstract: Implementations relate to determination and display of estimated hold durations for calls. In some implementations, a computer-implemented method includes obtaining an identifier of a target entity, the identifier usable by a first call device to initiate a call between the first call device and the target entity. A hold duration is determined that is an estimated amount of time before the call is matched to a human agent when the call is initiated at a particular time that is prospective to a current time. The hold duration is determined based on previous calls between one or more call devices and the target entity. The hold duration is provided for display by the first call device prior to the initiation of the call between the first call device and the target entity.

Abstract: The present disclosure describes a service which provides primary in-line deduplication. A streaming application program interface (API) may allow for streaming records into a storage system with high throughput and low latency. As part of this process, the API allows user to add identifiers as a field used for data deduplication. The deduplication service keeps a moving window of the identifiers in memory and does in-line deduplication by quickly determining whether data is a duplicate. Keeping only deduplication keys in memory reduces the cost of running the service. Moreover, the real-time nature of the moving window approach allows for storing deduplication information alongside the data and accessing it immediately on read. In this regard, read after write consistency is supported, and costs are reduced.

Abstract: This document describes low-latency Bluetooth connectivity in a wireless network in which a central node and a peripheral node establish a connection. During a first connection interval, the peripheral node receives a packet from the central node to synchronize communication with the central node, and based on receiving the packet, the peripheral node transmits a first fixed-length packet. If the first fixed-length packet fails to reach the central node, the peripheral node does not receive an acknowledgement, ACK, from the central node during the first connection interval and retransmits the first fixed-length packet during the first connection interval.

Abstract: A smart-home device may include a housing, a printed circuit board (PCB) inside the housing, an environmental sensor mounted to the PCB inside the housing, and a gasket that encloses the environmental sensor inside the housing to isolate the environmental sensor from an atmosphere inside of the housing while allowing an atmosphere outside of the housing to enter the gasket such that the environmental sensor can measure an aspect of the atmosphere outside of the housing.

Abstract: Methods, systems, and apparatus for training quantum evolutions using sub-logical controls. In one aspect, a method includes the actions of accessing quantum hardware, wherein the quantum hardware includes a quantum system comprising one or more multi-level quantum subsystems; one or more control devices that operate on the one or more multi-level quantum subsystems according to one or more respective control parameters that relate to a parameter of a physical environment in which the multi-level quantum subsystems are located; initializing the quantum system in an initial quantum state, wherein an initial set of control parameters form a parameterization that defines the initial quantum state; obtaining one or more quantum system observables and one or more target quantum states; and iteratively training until an occurrence of a completion event.

Abstract: A system and methods are disclosed for automatic focus detection with relative threshold-aware cell visibility for a scrolling cell collection. The method includes identifying a visible portion of a graphical user interface (GUI), wherein the visible portion of the GUI includes a plurality of cells, and determining a plurality of distance values, wherein each distance value of the plurality of distances values indicates a distance to a focus threshold in the GUI that a cell of the plurality of cells of the visible portion of the GUI has traveled since the visible portion became visible in the GUI, wherein the distance value comprises a relative threshold visibility value. The method further includes modifying a behavior of the visible portion of the GUI based on the distance value of each of the plurality of cells.

Abstract: A client device may receive a plurality of search result items. Each search result item may correspond to a media content item. The client device may present a user interface, including a scrollable search result area presenting a subset of the plurality of search result items, and a focus area presenting a preview of a media content item content item corresponding to one of the plurality of search result items. The client device may receive a user gesture in the scrollable search result area and in response, may determine whether the user gesture indicates a user interest in a second media content item. The client device may present a preview of the second media content item corresponding to the second search result item.