Abstract: A method for training hotword detection includes receiving a training input audio sequence including a sequence of input frames that define a hotword that initiates a wake-up process on a device. The method also includes feeding the training input audio sequence into an encoder and a decoder of a memorized neural network. Each of the encoder and the decoder of the memorized neural network include sequentially-stacked single value decomposition filter (SVDF) layers. The method further includes generating a logit at each of the encoder and the decoder based on the training input audio sequence. For each of the encoder and the decoder, the method includes smoothing each respective logit generated from the training input audio sequence, determining a max pooling loss from a probability distribution based on each respective logit, and optimizing the encoder and the decoder based on all max pooling losses associated with the training input audio sequence.

Abstract: Techniques are disclosed for applying a trained machine learning model to incoming voice communications to determine whether the voice communications are genuine or not genuine. The trained machine learning model may identify vocal attributes within the target call and use the identified attributes, and the training, determine whether the target call is genuine or not genuine. An applied trained machine learning model may include multiple different types of trained machine learning models, where each of different types of machine learning models is trained and/or configured for a different function within the analysis.

Abstract: Systems and methods of monitoring technology infrastructure using alerts indicative service events and tickets indicative of incidents reported to the support system, including transmitting, to a client via a network, structured support data including issue data and correlation data. The issue data represents issues, which are fewer than the number of tickets, generated by processing textual data of the tickets through a clustering engine implementing a generative probabilistic model and generating the correlation data by associating alerts and tickets by correlating alert-specific identifiers and ticket-specific identifiers. The identifiers are of least one of identifier times, locations, names, or descriptions. A prioritization engine is also disclosed.

Abstract: A system and method are disclosed capable of parsing a spoken utterance into a natural language request and a speech audio segment, where the natural language request directs the system to use the speech audio segment as a new wakeword. In response to this wakeword assignment directive, the system and method are further capable of immediately building a new wakeword spotter to activate the device upon matching the new wakeword in the input audio. Different approaches to promptly building a new wakeword spotter are described. Variations of wakeword assignment directives can make the new wakeword public or private. They can also add the new wakeword to earlier wakewords, or replace earlier wakewords.

Abstract: Implementations herein relate to information describing one or more internal states of a technical system. Implementations herein are provided for characterizing reliability of various different third party servers, at least when reporting third party device statuses, as well as adapting protocols for device ecosystems affected by such reliability. Latency can affect accuracy of device states represented by assistant devices. Certain servers can be characterized as especially delayed when reporting an updated device state in response to a user request, and, as a result, the third party server can be correlated to a metric that characterizes the relative latency of the third party server. When the metric fails to satisfy a particular threshold, a server and/or client associated with the “ecosystem” of third party devices can affirmatively operate to retrieve device state updates, rather than passively await updates from a corresponding third party server.

Abstract: Systems and methods for improved Spoken Language Understanding (“SLU”) are provided. The methods may comprise receiving an utterance from a user, contextualizing a plurality of words in the utterance, providing the contextualized words to the slot detector to determine the probability of a word forming the beginning or end of a slot to determine slots and nested slots, an intent classifier to determine the probability of a word conveying a user intent, and a slot classifier that applies specific labels to each slot and nest slot. The SLU method may employ a model and jointly trains the model for each task (determining beginning and end of slots, intents, and slot classifications) using a combined loss function.

Abstract: A method for detecting a hotword includes receiving a sequence of input frames that characterize streaming audio captured by a user device and generating a probability score indicating a presence of a hotword in the streaming audio using a memorized neural network. The network includes sequentially-stacked single value decomposition filter (SVDF) layers and each SVDF layer includes at least one neuron. Each neuron includes a respective memory component, a first stage configured to perform filtering on audio features of each input frame individually and output to the memory component, and a second stage configured to perform filtering on all the filtered audio features residing in the respective memory component. The method also includes determining whether the probability score satisfies a hotword detection threshold and initiating a wake-up process on the user device for processing additional terms.

Abstract: One or more computing devices, systems, and/or methods for detecting trigger phrases and transmitting electronic messages to devices are provided. For example, audio received via a microphone of a first device may be monitored. Responsive to detecting a first trigger phrase in a first audio segment identified during the monitoring, a first electronic message comprising instructions to activate a microphone function of a second device may be generated and the first electronic message may be transmitted to the second device. Responsive to detecting a second trigger phrase in a second audio segment identified during the monitoring, a second electronic message comprising instructions to activate a microphone function of a third device may be generated and the second electronic message may be transmitted to the third device.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for hotword detection on multiple devices are disclosed. In one aspect, a method includes the actions of receiving, by a first computing device, audio data that corresponds to an utterance. The actions further include determining a first value corresponding to a likelihood that the utterance includes a hotword. The actions further include receiving a second value corresponding to a likelihood that the utterance includes the hotword, the second value being determined by a second computing device. The actions further include comparing the first value and the second value. The actions further include based on comparing the first value to the second value, initiating speech recognition processing on the audio data.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, and a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution. The second attention-based encoder-decoder model is trained to classify output tokens based on input speech frames of a target speaker and simultaneously trained to maintain a similarity between the first output distribution and the second output distribution.

Abstract: Provided is a system and method for automated patient interaction. The method includes parsing a patient complaint comprising a plurality of words, determining a subset of patient queries from a plurality of patient queries based on the patient complaint and patient data, communicating the subset of patient queries to a first computing device; receiving, from the first computing device, responses to at least a portion of the subset of patient queries; generating output data based on the subset of patient queries and the responses; communicating the output data to a second computing device; receiving, from the second computing device, a user input corresponding to at least one patient query of the subset of patient queries; and training, based on the user input, at least one machine-learning algorithm configured to output at least one patient query based on at least one of the patient complaint and a subsequent patient complaint.

Abstract: A method and an electronic device for translating a speech signal between a first language and a second language with minimized translation delay by translating fewer than all words of the speech signal according to a level of understanding of the second language by a user that receives the translation.

Abstract: The present disclosure relates generally to determining intent based upon speech input using a dialog system. More particularly, techniques are described using matching-based machine learning techniques to identify an intent corresponding to speech input in a dialog system. These procedures do not require training when intents are added or removed from the set of possible intents.

Abstract: A speech separation model training method and apparatus, a computer-readable storage medium, and a computer device are provided, the method including: obtaining first audio and second audio, the first audio including target audio and having corresponding labeled audio, and the second audio including noise audio.

Abstract: This disclosure describes techniques for generating a conversation summary. The techniques may include processing at least one statement indication of the conversation to determine at least one statement that is a candidate highlight of the conversation. The techniques may further include applying linguistic filtering rules to the candidate highlight to determine the candidate highlight is an actual highlight. The techniques may further include generating the conversation summary including providing the actual highlight as at least a portion of the conversation summary.

Abstract: The systems and methods of the present disclosure generally relate to a data processing system that can identify and surface alternative requests when presented with ambiguous, unclear, or other requests to which a data processing system may not be able to respond. The data processing system can improve the efficiency of network transmissions to reduce network bandwidth usage and processor utilization by selecting alternative requests that are responsive to the intent of the original request.

Abstract: Implementations are directed to updating a trained voice bot that is deployed for conducting conversations on behalf of a third-party. A third-party developer can interact with a voice bot development system that enables the third-party developer to train, update, validate, and monitor performance of the trained voice bot. In various implementations, the trained voice bot can be updated by updating a corpus of training instances that was initially utilized to train the voice bot, and updating the trained voice bot based on the updated corpus. In some implementations, the corpus of training instances may be updated in response to identifying occurrence(s) of behavioral error(s) of the trained voice bot while the conversations are being conducted on behalf of the third-party. In additional or alternative implementations, the corpus of training instances may be updated in response to determining the trained voice bot does not include a desired behavior.

Abstract: A computer-implemented method includes receiving, by a computing device, a particular textual description of a scene. The method also includes applying a neural network for text-to-image generation to generate an output image rendition of the scene, the neural network having been trained to cause two image renditions associated with a same textual description to attract each other and two image renditions associated with different textual descriptions to repel each other based on mutual information between a plurality of corresponding pairs, wherein the plurality of corresponding pairs comprise an image-to-image pair and a text-to-image pair. The method further includes predicting the output image rendition of the scene.

Abstract: Systems and processes for operating an intelligent automated assistant are provided. An example method includes, at an electronic device having one or more processors and memory: receiving a natural language speech input; determining, based on the natural language speech input, a plurality of candidate intents; obtaining contextual data associated with the user device; ranking, based on the contextual data, the plurality of candidate intents using a machine learning model, wherein the machine learning model is pre-trained at least partially on the user device; determining a user intent based on the ranked candidate intents; and performing a task corresponding to the determined user intent.

Abstract: Techniques for voice-based interactions are described. In an example, a device presents a user interface on a display. The device starts an operational mode of the device. The operational mode restricts voice-based interactions with the user interface to a set of commands. The set of commands is defined in a language model that is stored on the device. Further, the device receives, at a microphone of the device, audio data corresponding to a natural language utterance and generates, from the audio data, text data that corresponds to the natural language utterance. The device determines, based at least in part on the language model, that semantics of the text data correspond to a command from the set of commands and presents, on the display, an outcome of performing the command.