Abstract: A method includes identifying a middlebox receiving network flow and communicating with one or more backend virtual machines. The method also includes receiving flow statistics corresponding to the network flow of the middlebox and determining whether the flow statistics satisfy an offload rule. The offload rule indicates when to migrate the network flow from the middlebox to an end host. When the flow statistics satisfy the offload rule, the method also includes migrating the network flow from the middlebox to the end host.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining, by a first sequence-training speech model, a first batch of training frames that represent speech features of first training utterances; obtaining, by the first sequence-training speech model, one or more first neural network parameters; determining, by the first sequence-training speech model, one or more optimized first neural network parameters based on (i) the first batch of training frames and (ii) the one or more first neural network parameters; obtaining, by a second sequence-training speech model, a second batch of training frames that represent speech features of second training utterances; obtaining one or more second neural network parameters; and determining, by the second sequence-training speech model, one or more optimized second neural network parameters based on (i) the second batch of training frames and (ii) the one or more second neural network parameters.

Abstract: This document describes a thermal-control system that is integrated into a security camera. The thermal-control system includes a combination of heatsinks and thermal interface materials with high thermal conductivities. The thermal-control system may transfer and spread energy from a high thermal-loading condition effectuated upon the security camera to concurrently maintain temperatures of multiple thermal zones on or within the security camera at or below prescribed temperature thresholds.

Abstract: A method includes obtaining first code in a high-level programming language. The first code represents a first function for performing one or more operations on plaintext. The method also includes converting the first code into an intermediate representation comprising a list of nodes. Each node of the list of nodes includes one or more logical operations. The method also includes converting the intermediate representation into a Boolean intermediate representation that includes a plurality of single-bit logical operations. The method also includes transpiling the Boolean intermediate representation into second code in the high-level programming language. The second code represents a second function for performing, using fully homomorphic encryption (FHE), the one or more operations on ciphertext.

Abstract: Various arrangements for reducing auditory spatial aliasing for a user are detailed herein. A first delay filter may be set that delays output of a first same audio signal by a first duration to a speaker of a first set of multiple speakers of a device compared to a second speaker of the first set of multiple speakers. A second delay filter may also be set that delays output of a second same audio signal by a second duration to a speaker of a second set of multiple speakers of the device compared to a second speaker of the second set of multiple speakers. The first same audio signal can be output using the first set of multiple speakers and the second same audio signal can be output using the second set of multiple speakers.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for allocating computing resources. In one aspect, a method includes receiving intent data specifying one or more computing services to be hosted by a computing network, requested characteristics of computing resources for use in hosting the computing service, and a priority value for each requested characteristic. A budget constraint is identified for each computing service. Available resources data is identified that specifies a set of available computing resources. A resource allocation problem for allocating computing resources for the one or more computing resources is generated based on the intent data, each budget constraint, and the available resources data. At least a portion of the set of computing resources is allocated for the one or more computing services based on results of evaluating the resource allocation problem to meet a particular resource allocation objective.

Abstract: In one embodiment, a system includes first host machines implementing a public-cloud computing environment, wherein at least one of the first host machines includes a resource manager that provides a public-cloud resource interface through which one or more public-cloud clients interact with one or more virtual machines, and second host machines implementing a private-cloud computing environment, wherein at least one of the second host machines includes one or more private-cloud virtual machines, wherein at least one of the first host machines further includes a private-cloud VM resource provider through which the resource manager interacts with the private-cloud virtual machines, wherein the VM resource provider translates requests to perform virtual machine operations from a public-cloud-resource interface to a private-cloud virtual machine interface, and the private-cloud virtual machines perform the requested virtual machine operations in response to receiving the translated requests from the VM resource

Abstract: This document describes techniques directed to active thermal-control of a floodlight and associated floodlights. As described, an example floodlight includes a first heat-transfer subsystem that uses a fully enclosed heat sink to transfer heat from an array of LEDs to a first housing component of the floodlight. The floodlight further includes a second heat-transfer subsystem to transfer heat from one or more PSUs to a second housing component of the floodlight. Described techniques include using thermistors located throughout the floodlight to actively monitor a temperature profile within the floodlight and, if one or more operating-temperature thresholds are violated, reducing power consumption within the floodlight.

Abstract: A method for configuring and deploying execution environments for software applications includes receiving a bootstrap execution environment from a data store, the bootstrap execution environment including a software application, the software application including a manifest of dependencies. The method includes executing the bootstrap execution environment and determining whether an enhanced execution environment is available from the data store. When the enhanced execution environment is available from the data store, the method includes receiving the enhanced execution environment from the data store and enhancing the bootstrap execution environment based on the received enhanced execution environment. When the enhanced execution environment is not available from the data, the method includes enhancing the bootstrap execution environment based on the manifest of dependencies to create the enhanced execution environment and storing the enhanced execution environment in the data store.

Abstract: A method for evaluating a verification model includes receiving a first and a second set of verification results where each verification result indicates whether a primary model or an alternative model verifies an identity of a user as a registered user. The method further includes identifying each verification result in the first and second sets that includes a performance metric. The method also includes determining a first score of the primary model based on a number of the verification results identified in the first set that includes the performance metric and determining a second score of the alternative model based on a number of the verification results identified in the second set that includes the performance metric. The method further includes determining whether a verification capability of the alternative model is better than a verification capability of the primary model based on the first score and the second score.

Abstract: This document describes techniques and systems that enable grid-based enrollment for face authentication. The techniques and systems include overlaying a three-dimensional (3D) tracking window over a preview image of the user's face displayed via a display device. The 3D tracking window includes a plurality of segments, which persist to correspond to an approximate direction that the user's face is facing. Based on the tracking, segments are highlighted to indicate the approximate direction that the user's face is facing, a camera captures enrollment images of the user's face facing that direction, and embeddings are generated based on the enrollment images and stored in a fixed grid of pose cells corresponding to various facial poses for use in face authentication. Responsive to generation and storage of the embeddings, an indication that the one or more segments are completed is provided.

Abstract: The various implementations described herein include a video camera assembly that includes: (1) a housing; (2) an image sensor encased in the housing and configured to capture activity of the smart home environment; (3) a wireless radio configured to transmit video frames captured by the image sensor to an electronic device via a remote server; (4) at least one infrared transmitter configured to selectively illuminate the smart home environment; (5) one or more circuit boards encased in the housing, the one or more circuit boards including at least one processor mounted thereon; and (6) a heating component coupled to the image sensor, the heating component configured to continuously maintain the image sensor at a temperature above a threshold temperature while the image sensor is capturing the activity of the smart home environment.

Abstract: Various arrangements for establishing a link between two electronic devices are presented herein. A pairing mode may be activated to establish the link. A first electronic wireless device can determine that a distance to a second wireless electronic device meets a distance threshold criterion. In response to the pairing mode being activated and the distance to the second wireless electronic device meeting the distance threshold criterion, a graphic element, such as a glow effect, can be presented indicative of direction and distance to the second wireless electronic device.

Abstract: Implementations described herein relate to providing device messages in image compilations based on user content. A computer-implemented method includes determining that a playback mode is entered on a device to display one or more image compilations, each image compilation including images for successive display. Message criteria are obtained for device messages, each device message indicating a device function associated with images of the image compilations or of a user image library. A device message is selected based on associated message criteria. An image is identified in the user image library, the image being in the image compilations, depicts a type of image feature in the image compilations, and/or has characteristic(s) modifiable by the device function. The device message is displayed during image compilation playback and includes the identified image and a message control option. In response to user selection of the message control option, the device function is initiated.

Abstract: Features described herein generally relate to enhancing functionality of electronic devices. Particularly, a notification is received from a wearable device that includes a position of the wearable device within an area, a determination is made that the wearable device is located within a predetermined distance from a controllable device, a graphical user interface page that includes buttons is sent to the wearable device, an indication that a button of the buttons has been activated is received from the wearable device, and a command signal that includes an instruction is sent to the controllable device. In this way, functionality of electronic devices can be extended with wearable devices.

Abstract: This document describes techniques and apparatuses directed at providing a tubular band with seamless lug integration for wearable devices. In aspects, the tubular band includes an extensible strap formed of interwoven fibers. The strap includes opposing outward-facing edges at opposing longitudinal ends. Each edge defines an opening within which a base section of a coupling device can reside. The coupling device further includes an attachment section, opposite the base section, which extends from the base section outwardly through the opening to engage a coupling connector of the wearable device. When the attachment section is connected to the coupling connector, the attachment section positions an end portion of the strap adjacent the opening to engage a surface of the wearable device. In this way, the strap can be coupled to a wearable device by positioning end portions of the strap against one or more surfaces of the wearable device, forming a seamless connection between the band and the wearable device.

Abstract: The present disclosure provides systems and methods that provides users with information pertaining to the audio properties at one or more points of interest. A database associated the audio properties with the points of interest is built using audio input received from devices at the points of interest. The device may determine that audio properties associated with the received audio input. The audio properties may determine a type of background noise and/or a volume of the background noise. If the type of background noise is music, the audio properties may further include a music genre, a title of a song, whether the music is recorded or there is a live band, etc. The audio properties associated with the point of interest may be updated in a database real time.

Abstract: Various arrangements for using a positioning tag to facilitate control of a smart home device are presented herein. A communication sessions can be established between a control device and a positioning tag via a first device-to-device communication protocol. In response to a trigger event indicative of a user desiring to control the separate smart home device being detected, a positioning enablement message can be transmitted to the positioning tag via the first device-to-device communication protocol. A signal used for positioning is transmitted by the positioning tag using a second device-to-device communication protocol. Based on the received signal via the second device-to-device communication protocol, an interface that controls the smart home device is activated.

Abstract: This document describes systems and techniques directed at expediting fingerprint authentication via variable refresh rate control and display luminance compensation. In aspects, a computing device having an under-display fingerprint sensor (UDFPS) and a touch-sensitive display includes a biometric authentication manager. Upon detecting at least one finger at or near the touch-sensitive display, the biometric authentication manager implements variable display refresh rates and selectively adjusts luminance settings for a high-luminance region of the touch-sensitive display for predetermined intervals. In so doing, at least one finger can be well-illuminated during UDFPS image capturing, facilitating UDFPS sensing and expediting fingerprint authentication.

Abstract: Methods, systems, and apparatus for automatic sharding and load balancing in a distributed data processing system. In one aspect, a method includes determining workload distribution for an application across worker computers and in response to determining a load balancing operation is required: selecting a first worker computer having a highest load measure relative to respective load measure of the other work computers; determining one or more move operations for a partition of data assigned to the first worker computer and a weight for each move operation; and selecting the move operation with a highest weight the selected move operation.

Abstract: A method for training a non-autoregressive TTS model includes obtaining a sequence representation of an encoded text sequence concatenated with a variational embedding. The method also includes using a duration model network to predict a phoneme duration for each phoneme represented by the encoded text sequence. Based on the predicted phoneme durations, the method also includes learning an interval representation and an auxiliary attention context representation. The method also includes upsampling, using the interval representation and the auxiliary attention context representation, the sequence representation into an upsampled output specifying a number of frames. The method also includes generating, based on the upsampled output, one or more predicted mel-frequency spectrogram sequences for the encoded text sequence.

Abstract: Techniques and apparatuses are described for adaptive phase-changing device power-saving operations. In aspects, a base station determines to transition an adaptive phase-changing device (APD) into an enabled APD-PS mode and determines an APD-PS configuration for the APD that specifies a framework for operating in the enabled APD-PS mode. The base station then directs the APD to operate in the enabled APD-PS mode by communicating the APD-PS configuration to the APD and transmits or receives wireless signals using a surface of the APD and based on the APD-PS configuration.

Abstract: The present disclosure provides for an arrangement of components within a hearable device, such as earbuds, to reduce H field coupling for E noise improvement. A battery is positioned above a magnet and speaker, such that the battery is not coaxially aligned with either the magnet or the speaker. An internal positive tab of a battery is positioned farthest from a speaker yoke, while an internal negative tab of the battery is closer to the speaker yoke. Magnetic flux poles of a contact magnet have different contributions for electric field coupling.

Abstract: An apparatus includes a processing device coupled to a memory storing instructions. The instructions may cause the processing device to receive photoplethysmography (PPG) data derived from signals associated with a PPG sensor oriented proximate a portion of an anatomy of a human subject and associated with a heart rate of the human subject; receive wb-accelerometer data derived from signals associated with a wb-accelerometer sensor oriented to sense the heart rate the human subject; and combine the PPG data and the wb-accelerometer data to generate a heart rate estimate.

Abstract: Systems and methods for adjusting light emitted from a display of a device are provided. The adjusting includes obtaining, from light of an environment detected by at least one sensor, a measured color of light of the environment, and obtaining, from light of the environment detected by at least one sensor, a measured brightness of light of the environment. In response to the obtaining the measured color and the measured brightness of light, a color of light emitted from the display is adjusted from an initial color to a target color. A brightness of light emitted from the display is adjusted from an initial brightness to a target brightness.

Abstract: Various arrangements for managing which earbud transmits microphone audio are presented herein. A first earbud of a pair of true wireless earbuds may receive a downstream audio packet addressed to a second earbud of the pair of true wireless earbuds. This first earbud can then transmit a negative acknowledgement indicating that the downstream audio packet was not properly received despite the first earbud having properly received the downstream audio packet. With the negative acknowledgment, the first earbud can then transmit an upstream audio packet using an identity of the second earbud.

Abstract: A method for disaggregated cache memory for efficiency in distributed databases includes receiving, from a user device, a first query requesting first data be written to a distributed database. The distributed database includes a plurality of nodes each controlling writes to a respective portion of the distributed database and a distributed cache pool caching a subset of the distributed database independently from the plurality of nodes. The method includes writing, using one of the plurality of nodes, the first data to the distributed database. The method also includes receiving, from the user device, a second query requesting second data be read from the distributed database. The method further includes retrieving, from the distributed cache pool, the second data. The method includes providing, to the user device, the second data retrieved from the distributed cache pool.

Abstract: A method includes receiving a source data set and a target data set and identifying a loss function for a deep learning model based on the source data set and the target data set. The loss function includes encoder weights, source classifier layer weights, target classifier layer weights, coefficients, and a policy weight. During a first phase of each of a plurality of learning iterations for a learning to transfer learn (L2TL) architecture, the method also includes: applying gradient decent-based optimization to learn the encoder weights, the source classifier layer weights, and the target classifier weights that minimize the loss function; and determining the coefficients by sampling actions of a policy model. During a second phase of each of the plurality of learning iterations, determining the policy weight that maximizes an evaluation metric.

Abstract: Features described herein generally relate to systems and methods for assembling an electronic device. Particularly, the electronic device can be assembled by holding a portion of the first component arranged in a first plane in a holding position with a stream of air output from a nozzle of an assembly system and aligning the portion of the first component with an opening in a portion of the second component arranged in a second plane parallel to the first plane while the portion of the first component is held in the holding position.

Abstract: A method includes receiving a scan request requesting to scan a set of network-connected assets designated for a network scan. For each respective network-connected asset, the method includes scanning, at a network security scanner using a first scanning privilege level, the respective network-connected asset. The method includes determining, based on the scan using the first scanning privilege level, whether the respective network-connected asset has a vulnerability. In response, the method includes scanning, at the network security scanner using a second scanning privilege level, the respective network-connected asset. The second scanning privilege level defines a lower level of access the network security scanner has than the first scanning privilege level. The method includes determining, based on the scans, an exposure level of the vulnerability. The method includes reporting the exposure level of the vulnerability to a user of the respective network-connected asset.

Abstract: This document generally describes systems and methods for dynamically adapting speech recognition for individual voice queries of a user using class-based language models. The method may include receiving a voice query from a user that includes audio data corresponding to an utterance of the user, and context data associated with the user. One or more class models are then generated that collectively identify a first set of terms determined based on the context data, and a respective class to which the respective term is assigned for each respective term in the first set of terms. A language model that includes a residual unigram may then be accessed and processed for each respective class to insert a respective class symbol at each instance of the residual unigram that occurs within the language model. A transcription of the utterance of the user is then generated using the modified language model.

Abstract: This document describes techniques and devices for a radar recognition-aided search. Through use of a radar-based recognition system, gestures made by, and physiological information about, persons can be determined. In the case of physiological information, the techniques can use this information to refine a search. For example, if a person requests a search for a coffee shop, the techniques may refine the search to coffee shops in the direction that the person is walking. In the case of a gesture, the techniques may refine or base a search solely on the gesture. Thus, a search for information about a store, car, or tree can be made responsive to a gesture pointing at the store, car, or tree with or without explicit entry of a search query.

Abstract: The various embodiments of a rotationally-locking magnetic alignment system and components thereof are described herein, which enable position locking and self-alignment along stepwise rotational increments. In aspects, a rotationally-locking magnetic alignment system can include magnetic alignment components, where each magnetic alignment component can include an array of magnetic field-inducing components that possess a particular magnetic polarity such that a first magnetic alignment component can attract, magnetically couple, and positionally lock a complementary second magnetic alignment component. In implementations, the array of magnetic field-inducing components can be arranged in one or more patterns, within the first magnetic alignment component and/or the second magnetic alignment component, allowing for the first magnetic alignment component to be positionally locked at any number of stepwise rotational increments.

Abstract: This document describes systems and techniques directed at compensating for voltage losses in organic light-emitting diode (OLED) displays. In aspects, a computing device having an OLED display and a luminance manager is configured to receive an indication of a luminance that is, or is intended to be, displayed by pixels of the OLED display. Responsive to and based on the received indication of luminance and a voltage loss, the luminance manager determines a luminance modification for the pixels of the OLED display. Based on the determined luminance modification, the luminance manager modifies the luminance that is displayed or modifies the luminance that is intended to be displayed by the pixels of the OLED display effective to compensate for the voltage loss.

Abstract: A method for predicting parametric vocoder parameter includes receiving a text utterance having one or more words, each word having one or more syllables, and each syllable having one or more phonemes. The method also includes receiving, as input to a vocoder model, prosodic features that represent an intended prosody for the text utterance and a linguistic specification. The prosodic features include a duration, pitch contour, and energy contour for the text utterance, while the linguistic specification includes sentence-level linguistic features, word-level linguistic features for each word, syllable-level linguistic features for each syllable, and phoneme-level linguistic features for each phoneme. The method also includes predicting vocoder parameters based on the prosodic features and the linguistic specification.

Abstract: A method for unstructured data analytics in data warehouses includes receiving an unstructured data query from a user, the unstructured data query requesting the data processing hardware determine one or more unstructured data files stored at a data repository that match query parameters. The method includes determining, using an object table, a set of unstructured data files stored at the data repository that matches the query parameters. The object table includes a plurality of rows, each row of the plurality of rows associated with a respective unstructured data file stored at the data repository, and a plurality of columns, each column of the plurality of columns comprising metadata associated with the respective unstructured data file of each row of the plurality of rows. The method includes returning, to the user, a structured data table including the determined set of unstructured data files.

Abstract: An apparatus includes a processing device coupled to a memory storing instructions. The instructions cause the processing device to receive photoplethysmography (PPG) data derived from signals associated with at least one PPG sensor; receive acoustic data derived from signals associated with at least one audio sensor oriented to sense a heart rate of a human subject; and combine the PPG data and the acoustic data to generate a heart rate estimate.

Abstract: A method for document extraction includes receiving, from a user device associated with a user, an annotated document that includes one or more fields. Each respective field of the one or more fields of the annotated document is labeled by a respective annotation. The method includes clustering, using a template matching algorithm, the annotated document into a cluster and inducing, using the annotated document, a document template for the cluster. The method includes receiving, from the user device, an unannotated document including the one or more fields. The method includes clustering, using the template matching algorithm, the unannotated document into the cluster and, in response to clustering the unannotated document into the cluster, extracting, using the document template, the one or more fields.

Abstract: A computer-implemented method for transcribing an utterance includes receiving, at a computing system, speech data that characterizes an utterance of a user. A first set of candidate transcriptions of the utterance can be generated using a static class-based language model that includes a plurality of classes that are each populated with class-based terms selected independently of the utterance or the user. The computing system can then determine whether the first set of candidate transcriptions includes class-based terms. Based on whether the first set of candidate transcriptions includes class-based terms, the computing system can determine whether to generate a dynamic class-based language model that includes at least one class that is populated with class-based terms selected based on a context associated with at least one of the utterance and the user.

Abstract: The various implementations described herein include methods, devices, and systems for monitoring activity in a home environment. In one aspect, a method performed at a voice-assistant device includes: detecting a sound; obtaining a determination as to whether the sound meets one or more monitoring criteria; and in accordance with a determination that the sound meets the one or more monitoring criteria, generating a notification.

Abstract: An electronic device receives a first plurality of images of a scene captured by an image sensor of an electronic device, receives an ambient light level proximate to the electronic device, and determines whether the ambient light level is less than a first threshold value. In accordance with a determination that the ambient light level is less than the first threshold value, the electronic device detects motion in the scene based on one or more of the first plurality of images. In accordance with detecting motion in the scene, the electronic device receives a second plurality of images of the scene captured by the image sensor of the electronic device, forms a composite image from two or more of the second plurality of images, and causes the composite image to be presented for display on a user device.

Abstract: A method for representing an intended prosody in synthesized speech includes receiving a text utterance having at least one word, and selecting an utterance embedding for the text utterance. Each word in the text utterance has at least one syllable and each syllable has at least one phoneme. The utterance embedding represents an intended prosody. For each syllable, using the selected utterance embedding, the method also includes: predicting a duration of the syllable by decoding a prosodic syllable embedding for the syllable based on attention by an attention mechanism to linguistic features of each phoneme of the syllable and generating a plurality of fixed-length predicted frames based on the predicted duration for the syllable.

Abstract: Techniques and apparatuses are described for integrated access backhaul with an adaptive phase-changing device (APD) are described. In aspects, a donor base station determines to include an APD in a communication path for the wireless backhaul link with a node base station and apportions APD access to the APD for the node base station. The donor base station then communicates with the node base station using the surface of the APD and based on the apportioned APD-access by using the surface to exchange wireless signals with the donor base station.

Abstract: A method of storing semi-structured data includes receiving user data from a user of a query system where the user data includes semi-structured user data. The method also includes receiving an indication that the semi-structured user data fails to include a fixed schema. In response to the indication that the semi-structured user data fails to include the fixed schema, the method further includes parsing the semi-structured user data into a plurality of data paths and extracting a data type associated with each respective data path of the plurality of data paths. The method additionally includes storing the semi-structured user data as a row entry in a table of a database in communication with the query system wherein each column value associated with the row entry corresponds to a respective one of the plurality of data paths and the data type associated with the respective data path.

Abstract: A pin fin placement assembly utilized to form pin fins in a thermal dissipating feature is provided. The pin fin placement assembly may place the pin fins on an IC die disposed in the IC package. The pin fin placement assembly may assist massively placing the pin fins with desired profiles and numbers on desired locations of the IC die. The plurality of pin fins is formed in a first plurality of apertures in the pin fin placement assembly. A thermal process is then performed to solder the plurality of pin fins on the IC die.

Abstract: This document describes methods implemented by and systems utilizing an alternating-current (AC) power-switching device. The AC power-switching device includes first and second current input/output (I/O) nodes, a current-limiting resistor, and a bypass switch connected between the first and second current I/O nodes. The bypass switch includes a first light-emitting-diode-input normally closed solid-state relay. Further, the AC power-switching device includes a rectifier, a voltage regulator, and a controller. The controller includes a programmable controller chip and is configured to detect a voltage level at the output of the rectifier, activate the bypass switch in response to the voltage level exceeding a first voltage level threshold, set a first lockout timer in response to activating the bypass switch, and deactivate the bypass switch in response to the voltage level dropping below a second voltage level threshold.

Abstract: This document generally describes systems, methods, devices, and other techniques related to speaker verification, including (i) training a neural network for a speaker verification model, (ii) enrolling users at a client device, and (iii) verifying identities of users based on characteristics of the users' voices. Some implementations include a computer-implemented method. The method can include receiving, at a computing device, data that characterizes an utterance of a user of the computing device. A speaker representation can be generated, at the computing device, for the utterance using a neural network on the computing device. The neural network can be trained based on a plurality of training samples that each: (i) include data that characterizes a first utterance and data that characterizes one or more second utterances, and (ii) are labeled as a matching speakers sample or a non-matching speakers sample.

Abstract: A method for training a streaming automatic speech recognition student model includes receiving a plurality of unlabeled student training utterances. The method also includes, for each unlabeled student training utterance, generating a transcription corresponding to the respective unlabeled student training utterance using a plurality of non-streaming automated speech recognition (ASR) teacher models. The method further includes distilling a streaming ASR student model from the plurality of non-streaming ASR teacher models by training the streaming ASR student model using the plurality of unlabeled student training utterances paired with the corresponding transcriptions generated by the plurality of non-streaming ASR teacher models.

Abstract: A method includes processing, using a speech recognizer, a first portion of audio data to generate a first lattice, and generating a first partial transcription for an utterance based on the first lattice. The method includes processing, using the recognizer, a second portion of the data to generate, based on the first lattice, a second lattice representing a plurality of partial speech recognition hypotheses for the utterance and a plurality of corresponding speech recognition scores. For each particular partial speech recognition hypothesis, the method includes generating a corresponding re-ranked score based on the corresponding speech recognition score and whether the particular partial speech recognition hypothesis shares a prefix with the first partial transcription.

Abstract: A device may identify a plurality of sources for outputs that the device is configured to provide. The plurality of sources may include at least one of a particular application in the device, an operating system of the device, a particular area within a display of the device, or a particular graphical user interface object. The device may also assign a set of distinct voices to respective sources of the plurality of sources. The device may also receive a request for speech output. The device may also select a particular source that is associated with the requested speech output. The device may also generate speech having particular voice characteristics of a particular voice assigned to the particular source.