Abstract: A computing system generates a plurality of training data sets for generating the NLP model. The computing system trains a teacher network to extract and classify tokens from a document. The training includes a pre-training stage where the teacher network is trained to classify generic data in the plurality of training data sets and a fine-tuning stage where the teacher network is trained to classify targeted data in the plurality of training data sets. The computing system trains a student network to extract and classify tokens from a document by distilling knowledge learned by the teacher network during the fine-tuning stage from the teacher network to the student network. The computing system outputs the NLP model based on the training. The computing system causes the NLP model to be deployed in a remote computing environment.

Abstract: Provided is a computer-implemented method for generating automatically annotations for tabular cell data of a table having column and rows, wherein the method includes: supplying raw cell data of cells of a row of the table as input to an embedding layer of a semantic type annotation neural network which transforms the received raw cell data of the cells of the supplied row into cell embedding vectors; processing the cell embedding vectors to calculate attentions among the cells of the respective row of the table encoding a context within the row output as cell context vectors; and processing the cell context vectors generated by the self-attention layer by a classification layer of the semantic type annotation neural network to predict semantic column type annotations and/or to predict relations between semantic column type annotations for the columns of the table.

Abstract: An electronic device includes an input device, a processor, and a memory The processor is configured to identify a first filter value of a first signal received from the input device. The processor is configured to receive a second signal after a first time elapses after the first signal is received. The processor is configured to receive a third signal after a second time elapses after the second signal is received. The processor is configured to compare a level of the second signal with a first threshold value for each of the at least one unit section of the second signal. The processor is configured to identify first information indicating that abnormal noise is present in a first section of the second signal. The processor is configured to perform filtering on the third signal based on the first filter value of the first signal according to the first information.

Abstract: A messaging system for audio character type swapping. Methods of audio character type swapping include receiving input audio data having a first characteristic and transforming the input audio data to an input image where the input image represents the frequencies and intensities of the audio. The methods further include processing the input image using a convolutional neural network (CNN) to generate an output image and transforming the output image to output audio data, the output audio data having a second characteristic. The input audio and output audio may include vocals. The first characteristics may indicate a male voice and the second characteristics may indicate a female voice. The CNN is trained together with another CNN that changes input audio having the second characteristic to audio having the first characteristic. The CNNs are trained using discriminator CNNs that determine whether audio has a first characteristic or a second characteristic.

Abstract: Techniques are described for improving the efficiency of dialog processing by prompting and processing user feedback substantially in real time. A dialog system receives speech input from a user and processes a first portion of the speech input to determine an initial discerned intent. The dialog system causes display of a visual indication of the initial discerned intent. The visual indication of the discerned intent is used to guide the dialog so that the user can correct or confirm the initial discerned intent in a natural fashion. If the initial discerned intent is inaccurate, the user can provide feedback correcting the dialog system, and the dialog system processes a second portion of the speech input to determine a modified discerned intent. Thus, the dialog system can use the feedback to correct misunderstandings on the fly.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer-readable storage media, for speech recognition using attention-based sequence-to-sequence models. In some implementations, audio data indicating acoustic characteristics of an utterance is received. A sequence of feature vectors indicative of the acoustic characteristics of the utterance is generated. The sequence of feature vectors is processed using a speech recognition model that has been trained using a loss function that uses a set of speech recognition hypothesis samples, the speech recognition model including an encoder, an attention module, and a decoder. The encoder and decoder each include one or more recurrent neural network layers. A sequence of output vectors representing distributions over a predetermined set of linguistic units is obtained. A transcription for the utterance is obtained based on the sequence of output vectors. Data indicating the transcription of the utterance is provided.

Abstract: A voice processing method includes: determining a historical voice frame corresponding to a target voice frame; determining a frequency-domain characteristic of the historical voice frame; invoking a network model to predict the frequency-domain characteristic of the historical voice frame, to obtain a parameter set of the target voice frame, the parameter set including a plurality of types of parameters, the network model including a plurality of neural networks (NNs), and a number of the types of the parameters in the parameter set being determined according to a number of the NNs; and reconstructing the target voice frame according to the parameter set.

Abstract: A computer extracts a vocal portion from a first audio content item and determines a first representative vector that corresponds to a vocal style of the first audio content item by applying a variational autoencoder (VAE) to the extracted vocal portion of the representation of the audio content item. The computer streams, to an electronic device, a second audio content item, selected from a plurality of audio content items, that has a second representative vector that corresponds to a vocal style of the second audio content item, wherein the second representative vector that corresponds the vocal style of the second audio content item meets similarity criteria with respect to the first representative vector that corresponds to the vocal style of the first audio content item.

Abstract: One example method involves operations for a processing device that include receiving, by a machine learning model trained to generate a search result, a search query for a text input. The machine learning model is trained by receiving pre-training data that includes multiple documents. Pre-training the machine learning model by generating, using an encoder, feature embeddings for each of the documents included in the pre-training data. The feature embeddings are generated by applying a masking function to visual and textual features in the documents. Training the machine learning model also includes generating, using the feature embeddings, output features for the documents by concatenating the feature embeddings and applying a non-linear mapping to the feature embeddings. Training the machine learning model further includes applying a linear classifier to the output features. Additionally, operations include generating, for display, a search result using the machine learning model based on the input.

Abstract: Named Entity Recognition (NER) in a user search query in a real-time search engine may be achieved by training a machine learning algorithm to create a trained model. The trained model may be configured to receive a user search query as input and to recognize and output zero or more named entities in the search query. The training may include an iterative training process in which further training data is added at each iteration, in some embodiments. The training may be based on three training data sets, in some embodiments. A first training data set may be based on user search and engagement activity. A second training data set may be artificially generated based on a catalog of named entity values. A third training data set may be based on optimized and supplemented data pairs sampled from the first training data set.

Abstract: A voice call method is provided. The method includes: obtaining a voice call state of a terminal system at a historical moment, the terminal system being provided with at least two audio acquisition devices; obtaining first voice signals acquired by the at least two audio acquisition devices at a current moment, and determining signal energy of the first voice signals; and determining a target audio acquisition device at the current moment from the at least two audio acquisition devices based on the voice call state at the historical moment and the signal energy of the first voice signals.

Abstract: A computer implemented method includes receiving information streams on a meeting server from a set of multiple distributed devices included in a meeting, receiving audio signals representative of speech by at least two users in at least two of the information streams, receiving at least one video signal of at least one user in the information streams, associating a specific user with speech in the received audio signals as a function of the received audio and video signals, and generating a transcript of the meeting with an indication of the specific user associated with the speech.

Abstract: Systems and methods for establishing user persona from audio interactions are disclosed, including a voice-based conversational AI platform having an acoustic analytical record engine and audio signal codification optimizer. The engine receives an audio sample indicative of voice conversation between an end user and a bot and transforms it into quantifiable and machine-ingestible power spectrum and acoustic indicators that uniquely represent the audio sample in the form of a feature vector. The optimizer ingests and processes the indicators to estimate likelihood of an attribute value representing the audio sample by constructing a convolutional neural network model for each attribute category. The optimizer establishes user persona attribute values across different attribute categories for the audio sample based on the estimated likelihood.

Abstract: Methods and systems for natural language processing/understanding of voice conversations are provided. Using natural language processing, a clinical condition is extracted from a voice conversation. A clinical ontology identifies clinical concepts associated with the clinical conditions. The clinical concepts are classified for documentation. The clinical concepts are searched and validated from within an individual's longitudinal record.

Abstract: Systems and methods for pre-training and fine-tuning of neural-network-based language models to reason directly over tables without generating logical forms. In some examples, a language model can be pre-trained using masked-language modeling tasks synthetically generated from tables pulled from a knowledge corpus. In some examples, the language model may be further pre-trained using pairs of counterfactual statements generated from those tables, and/or one or more statements that compare selected data from those tables. The language model may then be fine-tuned using examples that include only a question, an answer, and a table, allowing fine-tuning examples to be harvested directly from existing benchmark datasets or synthetically generated.

Abstract: An example operation may include one or more of transferring a copy of a plurality of revised translation data sets to be added to a software application into a grid structure, each revised translation data set comprising a prompt name in a first field, an interactive voice response (IVR) prompt in a second field, a translation of the IVR prompt into a different language in a third field, and a timestamp in a fourth field, identifying two revised translation data sets in the grid structure that comprise a duplicate prompt names in first fields thereof, deleting an oldest revised translation data set among the two identified translations data sets from the grid structure which has an oldest timestamp, and storing the grid structure without the deleted oldest revised translation data set in a repository.

Abstract: Methods and systems for natural language processing/understanding of voice conversations are provided. Using natural language processing, a clinical condition is extracted from a voice conversation. A clinical ontology identifies clinical concepts associated with the clinical conditions. The clinical concepts are classified for documentation. The clinical concepts are searched and validated from within an individual's longitudinal record.

Abstract: In some examples, data-driven social media analytics application synthesis may include generating, for each social media analytics application of a plurality of social media analytics applications, a corpus, performing term normalization, and generating a normalized corpus. An actor, an action and an object may be generated for each social media analytics application, which may be mapped into an embedding space. A semantic cohesion network may be generated for each social media analytics application, and a pair-wise semantic cohesion may be determined to identify semantically cohesive groups. A new social media analytics application may be synthesized based on the identified semantically cohesive groups.

Abstract: Disclosed is a speech processing apparatus and method using a densely connected hybrid neural network. The speech processing method includes inputting a time domain sample of N*1 dimension for an input speech into a densely connected hybrid network; passing the time domain sample through a plurality of dense blocks in a densely connected hybrid network; reshaping the time domain samples into M subframes by passing the time domain samples through the plurality of dense blocks; inputting the M subframes into gated recurrent unit (GRU) components of N/M-dimension; outputting clean speech from which noise is removed from the input speech by passing the M subframes through GRU components.

Abstract: Embodiments of the present disclosure relate to a method and an apparatus for speech interaction, and a computer readable storage medium. The method may include determining text information corresponding to a received speech signal. The method also includes obtaining label information of the text information by labeling elements in the text information. In addition, the method further includes determining first intention information of the text information based on the label information. The method further includes determining a semantic of the text information based on the first intention information and the label information.