Abstract: When creators generate media content in accordance with media programs, the media content is evaluated to identify any number of violations of policies, and to generate scores representing a level of risk that the creators will violate one or more of the policies in the future. Subsequently, media content of the creators is transmitted to listeners in accordance with the scores. In addition to audio data of creators or transcripts of the audio data, scores may be generated based on images associated with the creators, titles or summaries of media programs, or reports received from listeners. Scores calculated for creators may increase or decrease over time, depending on numbers of violations of policies by such creators, or other factors, and be utilized with a goal of protecting listeners against exposure to harmful content.

Abstract: Techniques for personalized batch and streaming speech-to-text transcription of audio reduce the error rate of automatic speech recognition (ASR) systems in transcribing rare and out-of-vocabulary words. The techniques achieve personalization of connectionist temporal classification (CT) models by using adaptive boosting to perform biasing at the level of sub-words. In addition to boosting, the techniques encompass a phone alignment network to bias sub-word predictions towards rare long-tail words and out-of-vocabulary words. A technical benefit of the techniques is that the accuracy of speech-to-text transcription of rare and out-of-vocabulary words in a custom vocabulary by automatic speech recognition (ASR) system can be improved without having to train the ASR system on the custom vocabulary. Instead, the techniques allow the same ASR system trained on a base vocabulary to realize the accuracy improvements for different custom vocabularies spanning different domains.

Abstract: Descriptors of machine learning tasks to be used to respond to analysis requests, indicating acceptable categories of runtime environments for the tasks and metrics to be collected from the tasks, are received via programmatic interfaces. In response to an analysis request, an orchestrator receives results from individual tasks as they become available, provides the results to other tasks, and causes a response to the request to be prepared using results from at least a subset of the tasks. Metrics collected from the tasks, and a visual representation of the tasks indicating their runtime environments are presented.

Abstract: Devices and techniques are generally described for voice-based content attribution for speech processing applications. In some examples, a request for voice-based content may be received from a first speech processing skill. First identifier data associated with the first speech processing skill may be received. A determination may be made that first content that is associated with the request. Voice-based output data describing the first content may be generated. A selection of the first content may be received. Attribution data may be determined based at least in part on the selection of the first content and the first identifier data. The attribution data may be sent to a remote computing device.

Abstract: An audio/video (A/V) device includes a first housing and a second housing rotatably coupled to the first housing. The first housing includes a camera oriented in a first direction, one or more microphones oriented in the first direction, one or more first lighting elements oriented in the first direction, one or more second lighting elements oriented in a second direction, one or more third lighting elements oriented in a third direction, and one or more first contacts. The second housing includes a loudspeaker oriented in a fourth direction and one or more second contacts configured to engage with the one or one first contacts. An engagement between the one or more first contacts and the one or more second contacts communicatively couples the loudspeaker to computing components with the first housing.

Abstract: Systems and methods are provided for handling file operations from a hosted computing instance via a secure compute layer. The secure compute layer is presented to the instance as a virtualized service device that is locally addressable by the instance. Software within the instance can submit file operations to the virtualized service device, which the secure compute layer can translate into calls to a network-accessible storage service. Results from the calls can then be passed back to the instance through the virtualized service device. As a result, the instance can communicate with a variety of different network services, without itself implementing network communications for those services.

Abstract: Techniques for detecting regressions with respect to the accuracy of an anomaly detection compute service in detecting anomalies in users' time series data. The techniques include providing an instrumented time series instrumented with a set of one or more anomalies to the anomaly detection service. The anomaly detection service detects a set of one or more anomalies in the instrumented time series. The precision and recall of the detected anomalies with respect to the instrumented anomalies is computed. From the computed precision and recall, an anomaly detection accuracy is computed as an F-score or F-measure. It is then determined whether a regression in anomaly detection accuracy has occurred by comparing the computed accuracy score to a threshold. If a regression has occurred, an alert can be generated or a recent change to the anomaly detection service can be rolled back.

Abstract: Provided herein are techniques for enabling communications between an access point (AP) and a number of stations (STAs) while maximizing transmission quality and/or minimizing power consumption. For example, an STA may be associated with a camera device used to capture video data and transmit the video data to the AP, and the AP may be configured to receive a registration request from the STA, register the STA, receive video data from the STA, and transmit the video data to a backend system such as a storage server. In some cases, the AP and STAs may communicate using a communication bandwidth that is sub-GHz but has a higher bandwidth relative to the low-bandwidth sub-GHz communication protocols discussed above. An example of such a high-bandwidth sub-GHz communication protocol is one that supports data transmission (e.g., video data transmission) using a bandwidth of 900 megahertz (MHz).

Abstract: Techniques and systems can obtain a first private key usable with a classical cryptography algorithm and a second private key usable with a post-quantum cryptography algorithm based on classical and post-quantum public keys hosted by a computer-implemented storage of an online service provider. A plurality of keys to perform a cryptography operation on data hosted by the computer-implemented storage can be generated, the plurality of keys generated based on at least the first and second private keys and a cryptography derivation function identified in the computer-implemented storage. The plurality of keys can be used to perform the cryptography operation on the data hosted by the computer-implemented storage.

Abstract: Users in a facility may move about and their authentication status may change accordingly. For example, upon entering a facility, a user may present identification data. At that point, the user is considered an authenticated user with persistent profile data. However, the identification of a user may not be presented at entry or may be unreliable, which causes the user to be considered an unauthenticated user with temporary profile data. In other scenarios, the identification of authenticated users may become unreliable based on their location or their proximity to other users. In cases where users are crowded at a location, the profile data for such users may be aggregated and may be based on common settings of the set of possible users in the group. A user interface may be presented to a user based on their designated profile data.

Abstract: When a query is received by a stateful data processing service, the service determines, for each table scan (and associated operations) of a query, whether to select the table scan for execution by a stateless data processing service. The selected table scans are sent to the stateless data processing service for execution, and results are received by the stateful data processing service. The stateful data processing service may also execute other table scans of the query locally, against a local data cache. If the data is not present in the local data cache, then the stateful data processing service will copy the table data into the local data cache before executing the table scan. A query result based on the remote and/or local table scans may then be returned to the client.

Abstract: Techniques for evaluating a natural language understanding (NLU) component and determining an action to resolve an issue processing a user input are described. The system determines which component is invoked by a baseline NLU component is processing the user input, and which component is invoked by an updated NLU component. Based on that information, the system selects the action to resolve the updated NLU component generating an undesired response to the user input.

Abstract: Generating instructions for programming a processing element array to implement a convolution operation can include determining that the convolution operation under-utilizes the processing element array. The convolution operation involves using the processing element array to perform a series of matrix multiplications between a set of filters and a set of input matrices. Each filter comprises a weight matrix. Each input matrix is assigned to a respective row in the processing element array. Under-utilization can be determined through detecting that less than a threshold number of rows would be used concurrently. In response to determining that the convolution operation under-utilizes the processing element array, instructions can be added for modifying the convolution operation to increase the number of rows used concurrently. The added instructions are executable to cause at least one input matrix to be processed in parallel across more rows compared to processing without modifying the convolution operation.

Abstract: Techniques for configuring device-specific skills as top-level skills are described. When a system receives a user input the system performs NLU processing to determine an intent of the user input. In some instances, the system may identify a device-specific skill associated with the device interacted with by the user. At least partially in parallel to performing NLU processing to determine the intent of the user input, the system may also perform NLU processing to determine a likelihood that the user input corresponds to an intent actionable by the device-specific skill. Once the system has finished NLU processing, the system may implement one or more prioritization rules to determine whether the user input should be sent to the device-specific skill or another skill of the system.

Abstract: Systems, methods, and computer-readable media are disclosed for modular robotic manipulation systems. In one embodiment, an example modular robot assembly may include a housing having a base, and a robotic manipulator coupled to the base, where the housing is configured to provide the robotic manipulator access to a first side, a second side, and a third side of the modular robot assembly. The module robot assembly may include a first camera system, a controller, and an optional display coupled to the housing. The modular robot assembly may be configured to be coupled to other modular robot assemblies, and the housing may be configured to be secured, such that the modular robot assembly can be independently transported.

Abstract: An input/output (I/O) device can initiate data migration of a virtual machine (VM) instance from a source device to a target device. The data migration of the VM instance may include migrating the data for the VM instance and tag data associated with the data. The data for the VM instance and the tag data may be stored together in a source memory. A first read request from the I/O device can enable a memory controller in the source device to read the data for the VM instance and the tag data together, store the tag data in a tag data buffer, and transmit the data for the VM instance to the target device. A second read request from the I/O device can read the stored tag data from the tag data buffer and transmit to the target device. The target device can write the data for the VM instance together with the tag data in the target memory.

Abstract: A system to manage database queries including storage devices to implement a data store to store database data and computing devices to implement a query engine. The query engine is configured to receive, from a client, a database query and initiate performance of the query at the data store. The query engine is configured to compare a performance time of the query with a performance time threshold. Based on the performance time exceeding the performance time threshold: send a query identifier for the query and a token indicating that the query has not been completed; and receive, from the client, an additional query comprising the token and the query identifier. Based on a determination that the performance time does not exceed the performance time threshold, send a response to the query to the client, the response comprising data requested by the query.

Abstract: A spatial search may be performed using representations of geometric shapes stored in a key-value store. A request to perform a spatial search may be received, the request including a geometric shape composed of one or more points. The points of the geometrical shape may be translated into one or more spatial indexes representing spatial cells using a space-filling curve. A key-value store may then be incrementally searched for each spatial index to identify spatial cells intersecting the geometric shape for which other known geometric shapes exist. The key-value store may then be searched to identify the known geometric shapes intersecting the geometric shape included in the search.

Abstract: Systems and methods to generate synthetic lip synchronization may generate source facial keypoints based on a source video, generate target facial keypoints based on a target audio, determine distances between the source facial keypoints and target facial keypoints, and transform or warp the source facial keypoints and associated surfaces to the target facial keypoints. In this manner, target video having synthetic lip synchronization that matches the target audio may be generated, and the target video may substantially preserve or maintain surface textures or features from the source video in the target video, thereby generating natural and believable synthetic lip synchronization corresponding to the target audio.