Abstract: Techniques are disclosed for migrating a computer application from an entity's premises to a web services platform. Data from multiple sources on the entity's premises is gathered and normalized into a common format. The normalized data is used to create a topology of the network on the entity's premises. This topology is analyzed to determine whether a computer application executing on the entity's premises may be migrated to the web service platform.

Abstract: Techniques are disclosed to implement an adaptive testing service (ATS) capable of automatically generating test cases for a network service to adapt test coverage to observed behaviors of the network service. In embodiments, the ATS uses telemetry data from a production version of the network service to identify classes of testable behaviors. Test cases are generated for the behaviors and assigned weights based on frequency or recency metrics of the behaviors. The test cases are stored in a test case repository, and may be used to monitor the production version of the network service or verify code changes to a development version of the network service. The test case weights may be used to select which test cases to run or determine whether code changes should be accepted or rejected. The test cases are evolved over time to adapt to behavior changes in the network service.

Abstract: Real-time audio enhancement for a target speaker may be performed. An embedding of a sample of speaker audio is created using a trained neural network that performs voice identification. The embedding is then concatenated with the input features of a trained machine learning model for audio enhancement. The audio enhancement model can recognize and enhance a target speaker's speech in a real-time implementation, as the embedding is in the same feature space of the audio enhancement model.

Abstract: An object data store may generate metadata responsive to a request that causes a scan of a data object for subsequent use in performing queries to the data object. A request may be received that causes a scan operation of the data object. As part of performing the scan one or multiple types of metadata describing the data object may be generated. The generated metadata may be applied to access the data object and perform a subsequently received query to the data object at the object data store.

Abstract: Maintaining the security of biometric data is an utmost priority. Biometric data is secured using one or more techniques. With one technique, biometric input such as images of a user's palm is used to generate first primary data (PD). The original biometric input is deleted from temporary secure storage while the first PD is securely stored. The first PD may then be processed later to determine a second PD. The first PD may then be deleted, and the second PD subsequently used. With another technique, biometric input or a PD may be processed by a first model to determine first secondary data (SD) that is representative of features of a particular user within a first embedding space. Later the PD may be processed by a second model to determine a second SD in a second embedding space. The first SD is deleted, and the second SD subsequently used.

Abstract: Techniques are described herein for generating a semantically related search query that differs by at least one lexical token from document data provided as input. A machine-learning model may be trained using supervised and/or deep learning algorithms and a training data set including historical queries and the document data identified as being semantically related to those queries. The semantically-related search query may be associated with the document and utilized in subsequent searches to match the document to a subsequent query based on identifying a lexical match between the subsequent search query and the semantically-related search query associated with the document. These techniques enable semantically related search queries to be assigned offline but utilized to identify semantically related documents using lexical matching techniques.

Abstract: During text-to-speech processing, a speech model creates output audio data, including speech, that corresponds to input text data that includes a representation of the speech. A spectrogram estimator estimates a frequency spectrogram of the speech; the corresponding frequency-spectrogram data is used to condition the speech model. A plurality of acoustic features corresponding to different segments of the input text data, such as phonemes, syllable-level features, and/or word-level features, may be separately encoded into context vectors; the spectrogram estimator uses these separate context vectors to create the frequency spectrogram.

Abstract: Systems and techniques for validation and generation of localized content for audio and video are described herein. The systems and techniques provide for training of twin neural networks to evaluate performance characteristics, sometimes referred to as content-auxiliary characteristics, of a localized performance. The localized performance may be validated or improved by identifying misalignment in the performance characteristics to ensure that localized content preserves content as well as creative intent and performance ability in the final product. The machine learning models trained using the techniques described herein may be used in connection with auto-localization processes to automatically generate high quality localized audio and video content.

Abstract: This disclosure describes a system for fulfilling items at a materials handling facility. In some instances, a predicted items list that identifies items that are likely to be picked by a user are determined and, when the user arrives at the materials handling facility, those predicted items are presented to the user for selection. For example, predicted items may be determined and an inventory holder that holds one or more of those predicted items may be retrieved by a mobile drive unit (such as a Kiva mobile drive unit) and presented to the user at a retrieval area. The user may pick the items they desire from the presented inventory holder.

Abstract: Media content may be mastered at a higher quality than is supported on various remote workstations to perform tasks with respect to that content, using transmission channels that do not support data transfer rates for large, high quality media content. A compressed version of this content may be transmitted over a first channel for use with various tasks on a remote workstation. For tasks such as color grading that benefit from this higher quality content, a separate but parallel communication channel is used to transmit a higher-quality version of this content. An uncompressed video stream can be encoded using a lossless codec to retain higher quality data. A high quality video stream can be transmitted over a separate transmission channel, and received to a decoder that can decode this stream to provide a high quality video signal for presentation via a grading monitor or other such high quality presentation device.

Abstract: Techniques for establishing connections between user devices and headless devices attempting to connect to networks. A headless device may attempt to connect to an access point that requires interaction with a captive portal webpage for access to a network. However, the headless device my lack a display to present the captive portal webpage. The headless device may establish a connection with a user device using a PAN protocol. The headless device may then receive the captive portal webpage received from the access point, and relay the webpage to the user device using the PAN protocol. A user may use the user device to interact with the captive portal webpage, and the user device may then relay interaction data back to the headless device using the PAN protocol. The headless device may then provide that interaction data to the access point to be provided access to the network.

Abstract: Techniques are described that enable users to configure the mirroring of network traffic sent to or received by computing resources associated with a virtual network of computing resources at a service provider network. The mirrored network traffic can be used for many different purposes including, for example, network traffic content inspection, forensic and threat analysis, network troubleshooting, data loss prevention, and the like. Users can configure such network traffic mirroring without the need to manually install and manage network capture agents or other such processes on each computing resource for which network traffic mirroring is desired. Users can cause mirrored network traffic to be stored at a storage service in the form of packet capture (or “pcap”) files, which can be used by any number of available out-of-band security and monitoring appliances including other user-specific monitoring tools and/or other services of the service provider network.

Abstract: Distributed orchestration of data retrieval for generative machine learning model may be performed. When a natural language request to perform a natural language task is received that is associated with a generative application, one or more data retrievers may be selected to access associated data repositories according to a previously specified retrieval configuration for the generative natural language application. The data may then be obtained by the selected data retrievers and used to generate a prompt to a generative machine learning model. A result of the generative machine learning model may then be used to provide a response to the natural language request to perform the natural language task.

Abstract: An index is created with split documents to retrieve and augment generation of a response to a natural language request using a generative machine learning model. When a natural language request is received, a search representation is generated and used to retrieve candidate portions of documents from the index. A relevancy ranking is performed to identify relevant portions of documents from the candidates and provide the relevant portions to prompt a generative machine learning model to provide a result for the natural language request.

Abstract: Intent classification is performed for executing a retrieval augmented generation pipeline for natural language tasks using a generative machine learning model. A natural language generative application with associated data repositories may submit a natural language task. A classification machine learning model is used to determine an intent for the natural language request. A number of iterations of a retrieval pipeline may be determined to perform the natural language task based on the intent. The natural language request may be processed through a retrieval pipeline according to the determined number of iterations before returning a result to the request.

Abstract: A set of alternative vocal input styles for specifying a parameter of a dialog-driven application is determined. During execution of the application, an audio prompt requesting input in one of the styles is presented. A value of the parameter is determined by applying a collection of analysis tools to vocal input obtained after the prompt is presented. A task of the application is initiated using the value.

Abstract: Template-based tuning is performed on a generative machine learning model where a shared template is used to tune the generative machine learning model across multiple natural language tasks. When a natural language request to perform a natural language task is received, portions of a shared template to complete are identified as part of generating a prompt. The generative machine learning model is instructed according to the generated prompt and a response to the request is returned based on a result of the generative machine learning model.

Abstract: Some embodiments provide a method for a circuit that executes a neural network including multiple nodes. The method loads a set of weight values for a node into a set of weight value buffers, a first set of bits of each input value of a set of input values for the node into a first set of input value buffers, and a second set of bits of each of the input values into a second set of input value buffers. The method computes a first dot product of the weight values and the first set of bits of each input value and a second dot product of the weight values and the second set of bits of each input value. The method shifts the second dot product by a particular number of bits and adds the first dot product with the bit-shifted second dot product to compute a dot product for the node.

Abstract: Techniques for handling natural language query processing are described. In some examples, a sequence-to-sequence model is used to handle a natural language query. Post-processing of a result of the sequence-to-sequence model utilizes fine-grained information from an entity linker. In some examples, the sequence-to-sequence model and aspects of a natural language query pipeline are used to handle a natural language query.