Abstract: Real-time low-complexity stereo speech enhancement with spatial cue preservation may be performed. A stereo speech enhancement system receives a stereo input signal (e.g., a left and right input signal). The stereo speech enhancement system estimates spatial cues for a target speaker and downmixes the stereo input signal into a monaural signal. A low-complexity model may then process the monaural signal to generate an enhanced monaural signal. The stereo speech enhancement system upmixes the enhanced monaural signal based on the estimated spatial cues for the target speaker, to generate an enhanced stereo output signal.

Abstract: Systems, methods, and computer-readable media are disclosed for systems and methods for on demand video recap generation. Example methods may include determining a first request to resume playback of video content, wherein the first request is associated with a user account, determining, using historical consumption data associated with the user account, a portion of the video content that has been consumed, where the portion of the video content comprises a first video segment and a second video segment, determining that a first video segment ranking of the first video segment satisfies an importance threshold, generating a video recap of the portion of the video content, wherein the video recap comprises the first video segment, and causing presentation of the video recap.

Abstract: Hyperparameters for tuning a machine learning system may be optimized using Bayesian optimization with constraints. The hyperparameter optimization may be performed for a received training set and received constraints. Respective probabilistic models for the machine learning system and constraint functions may be initialized, then hyperparameter optimization may include iteratively identifying respective values for hyperparameters using analysis of the respective models performed using an acquisition function implementing entropy search on the respective models, training the machine learning system using the identified values to determine measures of accuracy and constraint metrics, and updating the respective models using the determined measures.

Abstract: Low latency search for nearest neighbors in a dataset containing a large number of entries is improved using an error correction code (ECC) for partitioning data into clusters and retrieval. During initialization and preprocessing a d-dimensional space with clusters corresponding to ECC codewords is specified. Entries in the dataset are embedded into this space and associated with respective codewords, each codeword specifying a cluster. An index associates the codewords, clusters, and entries. During a query of the dataset, a query entry is processed to determine a query embedding in the d-dimensional space. The query embedding is used as input for a list decoder of the ECC. The list decoder provides a set of nearest codewords, with those codewords representing a set of candidate clusters that may contain nearest neighbors. The dataset entries associated with the candidate clusters may then be searched to determine query results comprising specific entries.

Abstract: Systems, methods, and computer-readable media are disclosed for annotating content data such as video data with annotation data (e.g., images, emoji, memes, stylized text, sounds) in near real time. Example methods may include determining transcribed text from the content data, associated annotation data with the transcribed text, and annotating the content data with some or all transcribed text and annotation data. Example methods may further include editing the annotated content data to generate modified annotated content data and sending the annotated content data and/or modified annotated content data to a device.

Abstract: Described are systems and methods for adjusting a takeoff platform for a vertical takeoff and landing (VTOL) aerial vehicle. The described systems and methods allow a VTOL aerial vehicle to be oriented prior to takeoff to reduce the effects that a VTOL aerial vehicle may experience due to weather conditions at takeoff. In view of the aerodynamic characteristics of a VTOL aerial vehicle and certain weather conditions, a heading, pitch, and/or roll of the VTOL can be provided to counteract the effects the VTOL may experience from the weather conditions at takeoff.

Abstract: System and methods are provided that can address dynamically determining capacity allocations for multiple services on the same computing devices in the same point of presences (PoPs). A routing and capacity management system receives logs from the services. The routing and capacity management system processes logs from the services to generate data that reflects resource usage at respective PoPs and/or border links. The routing and capacity management system dynamically determines capacity allocation for the services based on the generated data and instructs the services with the capacity allocation changes.

Abstract: An operational mode can be designated for a computing device, such as an integrated circuit. In particular, an input signal may be used to determine an operational mode, which may lead to bypassing or otherwise restricting one or more boot operations to permit use of the computing device. Such an approach provides improved flexibility to permit use of the computing device for testing or debugging while maintaining security features used during other operational modes.

Abstract: The coherency of data can be maintained even for accelerators that may not directly support, or be part of, a coherency protocol or domain. A pair of agents can be utilized that include an application accelerator agent and a coherent agent. The coherent agent can support the coherency protocol and track coherency information for data in the coherency environment, to ensure coherent access to the data. The application accelerator agent can work with the coherent agent to obtain the data according to the coherency protocol, then perform operations on the data corresponding to a respective application or process. Such an approach enables an accelerator to function like a processor or processor core that otherwise supports the coherency protocol and is part of the coherency domain. A single coherent agent can be used to maintain coherency information for multiple other components or agents.

Abstract: Techniques for performing conversation recovery of a system/user exchange are described. In response to determining that an action responsive to a user input cannot be performed, a system may determine a topic to recommend to a user. The topic may be unrelated to the original substance of the user input. The system may have access to various data representing a context in which a user provides an input to the system. The system may use these inputs and various data at runtime to make a determination regarding whether a user should be recommended a topic, as well as what that topic should be. The system may cause a question be output to the user, with the question asking the user about the topic, for example whether the user would like a song played, whether the user would like to hear information about a particular individual (e.g., artist), whether the user would like to know about a particular skill (e.g.

Abstract: The low power wide area network communication mechanism mainly allocates the semi-persistent scheduling (SPS) frames to the terminal devices having network access requirements sequentially by the gateway. The SPS cycle is composed of a plurality of SPS frames and at least one buffering frame. In each SPS cycle, a data reception window is activated by the gateway according to each allocated SPS frames; the SPS frames that the networked terminal devices belong to are activated synchronously within each SPS cycle in order for the networked terminal devices to perform a data transmission with the gateway during each SPS cycle and ensure that the data transmission does not overlap with the data sent by other terminal devices. The scheduling may be re-scheduled at a preset time interval to allocate the SPS frames of terminal devices that are inoperable or out of service area to the terminal devices having network access requirements.

Abstract: System and methods for IoT event detector correctness verification. Detector models (e.g., state-based models including variables, states, transitions and actions) take IoT device data as input and detect, based on the data, events that triggers actions. To verify a correctness of the models prior to deploying the models at scale, an event detector model correctness checker obtains a representation of a definition of the model, verifies, based on analysis of the model definition, whether the model complies with correctness properties, and generates a report indicating whether the model complies. Example correctness properties include a reachability correctness property that indicates that respective states or actions are reachable according to the definition of the event detector model. The analysis may be accessed via an interface element and may result in generation of a report that identifies a location of non-compliance within the model definition.

Abstract: A method for modifying recorded audio of a video title may include: identifying a video title available for presentation; identifying a portion of audio data, of the video title, representing dialogue of an actor of the video title; identifying a first utterance of the dialogue to be replaced in the video title; selecting, based on a replaceable amount of the audio data, a second utterance to replace the utterance word in the portion of the audio data; generating, based on a voice profile of the actor, first speech signals representing the actor uttering the second utterance; removing second speech signals representing the actor uttering the first utterance from the dialogue; adding the first speech signals into the dialogue using at least a portion of the replaceable amount; and generating a modified version of the video title comprising the dialogue with the speech signals added.

Abstract: A memory access tracking agent detects an access request directed to an in-memory representation of a target data structure. The agent obtains run-time context information pertaining to the request, including thread stack information. Indications of the occurrence of the access request and the context information are provided to a destination.

Abstract: A device, having executed a single sign-on operation, may maintain a private device key in volatile memory. A client application may request an authentication artifact as part of obtaining authentication credentials for the device. Responsive to the request, an artifact may be signed with the private device key to generate the requested authentication artifact. The single sign-on operation may be authenticated using an authentication device and a wrap key generated and applied by a secure cryptographic processor using an identifier based on an attestation certificate obtained from the authentication device after successful authentication. The wrap key may be used to decrypt the private device key stored on a persistent device storage.

Abstract: A computing device can obtain usage data associated with the device. The usage data indicate how the computing device operates or how the device is used. The device can analyze the usage data to recognize usage patterns. The usage patterns can correspond to recurring actions or tasks initiated by the user using the device, such as actions or tasks initiated due to the user's habits and/or routines. Based on the usage patterns, the device can determine a task that has a sufficient likelihood of being performed using the device within a specified or determined time (e.g., 5 minutes from now, one year from now, etc.). The device can provide information (e.g., recommendations) associated with the task, and likely relevant to the user. The user can use the provided information to perform the task, thereby increasing the ease of access or efficiency associated with performing the task.

Abstract: A method including establishing, at a first wireless audio device, a first communication channel with a second wireless audio device, receiving, at the first wireless audio device, audio data corresponding to a user utterance, determining that the audio data includes a wake word, transmitting at least a portion of the audio data from the first wireless audio device to a mobile device via a second communication channel, receiving an indication to establish a third communication channel with a second wireless audio device, and establishing the third communication channel with the second wireless audio device.

Abstract: A device can perform device arbitration, even when the device is unable to communicate with a remote system over a wide area network (e.g., the Internet). Upon detecting a wakeword in an utterance, the device can wait a period of time for data to arrive at the device, which, if received, indicates to the device that another speech interface device in the environment detected an utterance. If the device receives data prior to the period of time lapsing, the device can determine the earliest-occurring wakeword based on multiple wakeword occurrence times, and may designate whichever device that detected the wakeword first as the designated device to perform an action with respect to the user speech. To account for differences in sound capture latency between speech interface devices, a pre-calculated time offset value can be applied to wakeword occurrence time(s) during device arbitration.

Abstract: A service provider manages access control to multiple services through an authentication system. One or more services are able to fulfill requests at least in part by submitting requests to other services of the service provider. Such a service is able to obtain, from the authentication system, information that can be passed on to one or more other services to enable the one or more other services to determine request validity without having to contact the authentication system. The information may include, for example, one or more responses that the one or more other services would have received had the one or more services contacted the authentication system themselves.

Abstract: A message indicating an auxiliary task associated with traffic transmitted via a virtual router between a pair of isolated networks is received at an offloading device. A stack multiplexer at the offloading device selects a protocol stack instance to process the message. A result of the auxiliary task is obtained by the multiplexer from the selected protocol stack instance and transmitted to the virtual router, where it is used to transmit a packet between the isolated networks.

Abstract: Data about concentration of one or more types of molecules present within a human body are determined noninvasively using radio frequency (RF) signals. Signals at several different frequencies at very low power levels are emitted using an antenna mounted to a wearable device. Operating values, such as changes to the impedance of the antenna resulting from interaction between the RF signals and the user, are associated with the concentration of one or more types of molecules within the user. These changes in impedance may be relatively small. Circuitry in the device determines internal impedance of the circuitry, compensating for internal impedances and improving overall accuracy of the measured impedance resulting from interaction with the user.

Abstract: Systems and methods for determining semantic relationships in digital content are provided. Digital content may be processed to detect and extract one or more entities present in at least a subset of the digital content. The entities may include at least one image and at least one body of text. One or more aspects of the entities may be analyzed to filter out a subset of the entities based on a determined importance of the entities to the subset of the digital content. An anchor entity may be determined from one or more remaining entities after the filtering. Relationships between the one or more remaining entities and the anchor entity may be determined, and a presentation pattern for presenting the related entities may be determined. The presentation pattern may be stored to a repository and used as training data to detect future relationships within digital content.

Abstract: Devices and techniques are generally described for an edge computing system. In various examples, the edge computing system may comprise a host kernel comprising a kernel-based virtual machine hypervisor. In some examples, the edge computing system may comprise virtualization software effective to communicate with the kernel-based virtual machine hypervisor to execute guest virtual machines. In various further examples, the edge computing system may comprise an engine virtual machine with access to at least one hardware component. The edge computing system may further comprise a control plane virtual machine. The control plane virtual machine may include components effective to receive a first request and determine an application corresponding to the first request, and a virtual machine manager effective to control the virtualization software to generate a virtual machine, configured in accordance with a configuration specific for the virtual machine, for executing the application.

Abstract: Techniques for facilitating natural language interactions with visual interactive content are described. During a build time, a system analyzes various websites and applications relating to a particular user goal to understand website and application navigation and information relating to the user goal. The learned information is used to store configuration data. During runtime, when a user request performance of an action, the system engages in a dialog with the user to complete the user's goal. The system uses the stored configuration data to determine actions to be performed at a website or application to complete the user's goal, and determines system responses to present to the user to facilitate completion of the goal. Such system responses may request information from the user, may inform the user of information displayed at the website or application, etc.

Abstract: Systems and methods for device-to-account anomaly detection are disclosed. For example, device information associated with user account data is queried from a user registry and from an external system. The device information from both sources is hashed or otherwise formatted and compared to determine whether a dissimilarity is present. A difference value may be generated and utilized to determine whether an anomaly is present for device-to-account associations as between the user registry and the external system. Utilizing the difference value and/or historical difference values associated with the external system, one or more actions for correcting the anomaly may be selected and performed.

Abstract: Conversion to a virtual target may be performed for a source of a data store migration. A request that selects one of multiple data store conversion targets as a virtual target of data store migration from a source of the data store migration may be received. One or more conversion scripts for the data store migration may be generated based on an analysis of a possible conversion between the source of the data store migration and the virtual target of the data store migration using metadata obtained from the source of the data store migration that describes the source of the data store migration and a model corresponding to the virtual target of the data store migration. The conversion scripts may be stored for later execution.

Abstract: A garage door is controlled to open using an audio/video (A/V) recording and communication device. The device detects a vehicle within an area about the garage door and receives, using a camera, image data representative of an object associated with the vehicle. The device also compares the image data representative of the object associated with the vehicle with previously stored image data and identifies the object based on the comparing of the image data representative of the object with the previously stored image data. The device also authenticates an electronic device within the vehicle that is associated with the object by wirelessly communicating with the electronic device and determines, based at least in part on the authenticating of the electronic device within the vehicle, that the vehicle is an authorized vehicle. The device also transmits an actuation command to the garage door to cause the garage door to open.

Abstract: An apparatus comprising an array of electrical conductors connected to circuitry that determines data indicative of objects at particular locations based on capacitance at junctions of conductors in the array. The apparatus comprises a first set of conductors arranged on a first substrate and a second set of conductors arranged on a second substrate to form an array. The first and second substrates are separated by an air gap within which the electrical conductors are positioned. Each conductor may be independently addressable, having a proximal end connected to readout electronics and a distal end that is not. The apparatus may be cut, at the ends distal to where the conductors connect to the circuitry, to fit a desired location, such as on a store shelf, while remaining operational.

Abstract: A single instruction multiple data (SIMD) processor is used to implement a dropout layer between a first layer and a second layer of a neural network. The SIMD processor can implement the dropout layer by setting one or more elements in an output tensor of the first layer to zero before providing it as an input tensor to the second layer. Setting of the one or more elements to zero is based on a dropout rate, and pseudo-random numbers generated by a random number generator in the SIMD processor.

Abstract: A technology is described for correlating application performance to external event types. An example method may include obtaining electronic content from a content server associated with an electronic content provider and extracting a potential reference to an external event from the electronic content, where the external event may impact performance of an application hosted on a computing resources. The reference may be identified as being associated with an external event type using a categorization technique that associates the reference in the electronic content to one or more external event types, and an external event log may be generated to include event data for the external event type to allow correlation of the external event type to performance of the application hosted on the computing resources.

Abstract: Methods for simplifying a dependency graph in a neural network accelerator are provided. Computations and data movements for the neural network accelerator may be described with a flow graph, where graph nodes represent computation or data movement operations and graph edges represent dependencies between operations. A flow graph may contain redundant edges that can be removed while retaining the reachability of each of the nodes in the graph. To identify redundant edges, a compiler may generate vector clocks to track the relationships of operations performed by various execution engines prior to execution of a program reaching a given node or operation. Redundant edges may be identified and removed based on the relative values of the vector clocks to reduce the complexity of the graph.

Abstract: Devices and techniques are generally described for three dimensional room modeling. In various examples, a three-dimensional (3D) room model comprising at least a first wall and a floor may be received. A first sphere sized and shaped such that the 3D room model fits within the first sphere may be determined and a virtual camera may be positioned on the first sphere at a first position. A first command may be received to move the virtual camera in a first direction. The virtual camera may be translated in the first direction along a surface of the first sphere to a second position. A view of the interior of the 3D room model may be displayed from a second viewpoint of the virtual camera at the second position.

Abstract: Systems, devices, and methods are provided for stable pseudonymous identifiers. A system may obtain first data, and obtain a first cryptographic key from a cryptography service. The system may determine, based at least in part on the first data and the first cryptographic key, a first hash output of a hash function. The system may determine that a mapping table does not include a first record associated with the first hash output. The system may create, in the mapping table, a first record comprising the first hash output and a random value. The system may provide the random value as a stable pseudonymous identifier of the first data. If the first cryptographic key is subject to a key rotation, the pseudonymous identifier of the first data may remain constant even though the value of the key changes as a result of the key rotation.

Abstract: A system can be configured to receive and analyze an image captured by a customer at a delivery location to enable identification and verification of the delivery location upon arrival of a deliverer at the delivery location. The system can utilize algorithms trained to determine identifying features of the delivery location from the image and/or provide the image to the deliverer for manual verification. The system can provide the deliverer with the image and additional information specific to the delivery location based on one or more geofences associated with the delivery location. The image, the additional information and the manner of presentation can be configured to guide the deliverer to approved drop-off locations and ensure package placement for delivery to a customer.

Abstract: A real-time data validation system prevents problematic updates from being replicated from a primary database to a replica database. The data validation system may inject a time delay into the replication process, so that when an update in a replication stream fails a specified data validation, the data validation system will prevent the update from being applied to the replica. The replication process may be halted to take snapshot of the replica database during the halt, and resumed after the snapshot. Accordingly, if the problematic update is later confirmed to be invalid, the replica database can be rolled back to a point before the problematic update. Advantageously, the disclosed system can be used to implement a variety of data validations on top of a database replication process to protect the replica database from receiving undesired updates.

Abstract: A chat input identifier may receive various chat inputs based on voice or text inputs from a user. The chat input identifier may apply different filters to the chat inputs to identify one or more chat correction pairs (e.g., chat input with errors, corrected chat input) from among the plurality of chat inputs. The chat correction pairs are used to train an auto-correction model. The trained auto-correction model receives a given chat input that has one or more errors. The auto-correction model processes the given chat input to generate a corrected version of the given chat input (without the need to obtain a correction from the user). The corrected chat input is then provided to a dialog-driven application.

Abstract: A supplemental content retrieval system receives requests for supplemental content from various content providers, and makes respective requests to supplemental content provider(s). Responses to the requests may be subject to a time limit. The requests to the supplemental content provider(s) may be made in parallel or serially, and may sometimes include information based on responses from another content provider. The supplemental content retrieval system receives responses with content items from the supplemental content providers and, for items in the responses that do not specify a category, determines categories of the items (e.g., based on item metadata, interactive links in the content item, item identifiers for the items, etc.). The supplemental content retrieval system applies a constraint (e.g., an inter-item category constraint or similar, which also may be associated with a time limit) to the categorized items to prepare responses to respective content provider requests.

Abstract: Queries may be accelerated with compound disjunctive conditions. A query is received and a compound disjunctive condition identified. Different strategies, such as inferring and adding additional disjunctive join conditions, may be implemented as part of query planning. These additional disjunctive join conditions may be included in a query plan generated for the query, which is then executed to return a result for the query.

Abstract: Techniques for orchestration of directory management updates across regions of a provider network are described. A method for orchestration of directory management updates across regions may include receiving a request at a service in a home region of a provider network to perform a cross-region service update, executing a cross-region workflow corresponding to the cross-region directory service update, and updating one or more resources of the service in each of a plurality of regions of the provider network based on the cross-region workflow.

Abstract: Systems and methods are described for generating a schedule for resources, such as agents for a contact center service, and mores specifically for decomposing the scheduling problem into groups. In some aspects, a request to generate a schedule spanning a first time period for agents may be received by a scheduling service. Agents having at least one same shift characteristic may be grouped into a scheduling group such that the agents in a given scheduling group are interchangeable. Portions of the workload may be assigned to the different scheduling groups. A set of empty shifts may be produced for each of the scheduling groups to process the corresponding portion of the workload. Shift activities may be assigned to the set of empty shifts for each scheduling group to generate unrostered shifts, the unrostered shifts making up the schedule.

Abstract: Systems, methods, and computer-readable media are disclosed for passive throttling of bulk material on conveyor systems. In one embodiment, an example system may include a first conveyor having a first width, a second conveyor having a second width, the second conveyor disposed adjacent to the first conveyor, a third conveyor disposed adjacent to the second conveyor, a chute, and a fourth conveyor. Items falling from the second conveyor may be guided to the fourth conveyor via the chute. The first conveyor, the second conveyor, and the third conveyor may move in a first direction, and the fourth conveyor may move in a second direction.

Abstract: Described are systems and methods for providing a fault tolerant electrical power system that is configured to continue operating through various fault and/or failure conditions, so as to withstand certain single points of fault and/or failure conditions. Exemplary embodiments of the present disclosure can provide an exemplary electrical power system that includes various components that are electrically coupled to a shared direct current (DC) bus via protection devices that are configured to isolate the various components in the event of certain fault and/or failure conditions. Additionally, the exemplary electrical power system can also include a high-impedance midpoint grounding connection configured to shift the voltages of the shared DC bus in view of certain fault and/or failure conditions, as well as introducing certain separations (e.g., physical and/or electrical) to mitigate the possibility of certain faults and/or failure conditions.

Abstract: Systems and methods for device-state quality analysis include storing data indicating device state of a device and receiving additional data indicating the device state of the device. Data is then generated based on the stored and received device state data that indicate that the stored data is inaccurate. Confidence value data is then generated utilizing the generated data and prior device-state accuracy data.

Abstract: A system, method, and computer-readable medium for managing content delivery networks are provided. A content delivery network can train a set of machine learned algorithms corresponding to anomaly detection in the hosting of content. The content delivery network can obtain log information from a plurality of POPs and additional information for use as inputs to selected machine learned algorithms. The content delivery network can the generate a set of recommendations as either CDN performance recommendations or CDN analysis recommendations for customers.

Abstract: Systems and methods are disclosed for container dumpers having controlled unloading. In one embodiment, an example container unloading system may include a container support mechanism configured to secure a container, where the container has items inside, and a first conveyor configured to rotate from a first position to a second position. The first conveyor may be disposed in a vertical orientation adjacent to a sidewall of the container in the first position, and the first conveyor may be disposed in a non-vertical orientation adjacent to the sidewall of the container in the second position. The system may include a controller configured to cause the container support mechanism to rotate the container from a vertical orientation to a rotated orientation, and cause the first conveyor to be actuated, such that the items are conveyed out of the container by the first conveyor.

Abstract: This disclosure proposes systems and methods for entity resolution using speech recognition data. The system can receive audio data representing an utterance and perform automatic speech recognition (ASR) processing on the audio data to generate at least a first ASR data and a second ASR data. The system can perform natural language understanding (NLU) processing on the first ASR data to determine intent data and an indication of an entity. The system can determine a first portion of the first ASR data that corresponds to the indication of the entity. The system can determine a second portion of the second ASR data that corresponds to the indication of the entity without performing NLU on the second ASR data. The system can perform entity resolution (ER) on the second portion to identify a first entity.

Abstract: Techniques for generating synthetic data are described. An exemplary approach includes receiving one or more requests to generate synthetic data based on a first dataset; generating the synthetic dataset is generated according to the request by choosing a set of synthetic datapoints between pairs of datapoints of the first dataset along a line connecting them while sampling a likely value of a local probability distribution; and providing the synthetic dataset as configured by the request.

Abstract: Techniques to improve the operation of homomorphic circuits. Systems and methods described herein may improve the operation of a homomorphic circuit by analyzing a circuit to reduce the multiplicative depth of the circuit. Techniques described herein may be utilized to perform linear algebra operations with mixed encoding units, according to a homomorphic encryption scheme.