Abstract: A server is provided. The server includes a communication circuitry, and at least one processor operatively connected with the communication circuitry. The at least one processor may be configured to, in response to traffic of a plurality of speeches to wake up a voice assistant feature, received within a preset period being a preset value or more, generate a plurality of clusters based on similarities between the plurality of speeches, and determine whether to respond to each of speeches included in each of the plurality of clusters based on similarities between the speeches included in each of the plurality of clusters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for language models using domain-specific model components. In some implementations, context data for an utterance is obtained. A domain-specific model component is selected from among multiple domain-specific model components of a language model based on the non-linguistic context of the utterance. A score for a candidate transcription for the utterance is generated using the selected domain-specific model component and a baseline model component of the language model that is domain-independent. A transcription for the utterance is determined using the score the transcription is provided as output of an automated speech recognition system.

Abstract: Provided is a pattern recognition apparatus to provide classification robustness to any kind of domain variability. The pattern recognition apparatus 500 based on Neural Network (NN) includes: NN training unit 501 that trains an NN model to generate NN parameters, based on at least one first feature vector and at least one domain vector indicating one of subsets in a specific domain, wherein, the first feature vector is extracted from each of the subsets, the domain vector indicates an identifier corresponding to the each of the subsets; and NN verification unit 502 that verifies a pair of second feature vectors in the specific domain to output whether the pair indicates same individual or not, based on a target domain vector and the NN parameters.

Abstract: Systems and methods are disclosed for generating internal state representations of a neural network during processing and using the internal state representations for classification or search. In some embodiments, the internal state representations are generated from the output activation functions of a subset of nodes of the neural network. The internal state representations may be used for classification by training a classification model using internal state representations and corresponding classifications. The internal state representations may be used for search, by producing a search feature from an search input and comparing the search feature with one or more feature representations to find the feature representation with the highest degree of similarity.

Abstract: Techniques are provided for training, by a system operatively coupled to a processor, an attention weighted recurrent neural network encoder-decoder (AWRNNED) using an iterative process based on one or more paragraphs of agent sentences from respective transcripts of one or more conversations between one or more agents and one or more customers, and based on one or more customer response sentences from the respective transcripts, and generating, by the system, one or more groups respectively comprising one or more agent sentences and one or more customer response sentences selected based on attention weights of the AWRNNED.

Abstract: A pattern recognition apparatus for discriminative training includes: a similarity calculator that calculates similarities among training data; a statistics calculator that calculates statistics from the similarities in accordance with current labels for the training data; and a discriminative probabilistic linear discriminant analysis (PLDA) trainer that receives the training data, the statistics of the training data, the current labels and PLDA parameters, and updates the PLDA parameters and the labels of the training data.

Abstract: A computer-implemented method is presented for detecting anomalies in dynamic datasets generated in a cloud computing environment. The method includes monitoring a plurality of cloud servers receiving a plurality of data points, employing a two-level clustering training module to generate micro-clusters from the plurality of data points, each of the micro-clusters representing a set of original data from the plurality of data points, employing a detecting module to detect normal data points, abnormal data points, and unknown data points from the plurality of data points via a detection model, employing an evolving module using a different evolving mechanism for each of the normal, abnormal, and unknown data points to evolve the detection model, and generating a system report displayed on a user interface, the system report summarizing the micro-cluster information.

Abstract: Systems and methods for providing customized automatic speech recognition (ASR) in a customer support system are disclosed. In an example method, one or more data sources for training an ASR language model associated with the customer support system are identified, and one or more weighting models are selected, each weighting model applying a corresponding weight to each data source of the one or more data sources. The ASR language model is then trained based at least in part on the one or more data sources and the one or more weighting models, and a transcript may be generated for one or more customer support calls of the customer support system using the trained ASR language model.

Abstract: A dynamic recognition method includes, when the terminal device detects that the user is in a first distance range, obtaining, by the terminal device, first feature information of the user. The method further includes performing first identity authentication on the first feature information of the user, where the first feature information includes facial feature information, voice feature information, or behavioral feature information. The method further includes increasing, by the terminal device, a level of a default threshold of second identity authentication when the first identity authentication succeeds. The method further includes, when the terminal device detects that the user is in a second distance range, obtaining, by the terminal device, second feature information of the user, and performing second identity authentication on the second feature information of the user based on the default threshold whose level is increased.

Abstract: A system for ontology-aware sound classification. The system includes an electronic processor that is configured to create a first graph based on relationships between fine audio classification labels and create a second graph based on relationships between coarse audio classification labels. The electronic processor is also configured to receive an audio clip including one or more sounds, execute a first graph convolutional network with the first graph as input, and execute a second graph convolutional network with the second graph as input. Using the outputs of the first graph convolutional network and the second graph convolutional network, the electronic processor is configured to determine one or more coarse labels, one or more fine labels, or both to classify the one or more sounds in the audio clip.

Abstract: A Cyber-Physical System (“CPS”) may have monitoring nodes that generate a series of current monitoring node values representing current operation of the CPS. A normal space data source may store, for each monitoring node, a series of normal monitoring node values representing normal operation of the CPS. An abnormal data generation platform may utilize information in the normal space data source and a generative model to create generated abnormal to represent abnormal operation of the CPS. An abnormality detection model creation computer may receive the normal monitoring node values (and generate normal feature vectors) and automatically calculate and output an abnormality detection model including information about a decision boundary created via supervised learning based on the normal feature vectors and the generated abnormal data.

Abstract: Techniques for identifying vocabulary associated with semantic objects used for generating natural language text with a natural language generation (NLG) system, the semantic objects including a first semantic object having a first set of ordered attributes. The techniques include: obtaining text segments; identifying, from among the text segments and using at least one first machine learning classifier, groups of text segments corresponding to respective semantic objects in the plurality of semantic objects; identifying, from the groups of text segments and using at least one second machine learning classifier, a plurality of vocabularies for the plurality of semantic objects; and generating natural language text using the NLG system, the plurality of vocabularies, and the plurality of semantic objects; and outputting the generated natural language text.

Abstract: One embodiment provides a method, including: obtaining historical information for equipment having at least one control, wherein the historical information indicates a setting for the at least one control during operation of the equipment and identifies operating performance of the equipment corresponding to the indicated setting; receiving a goal for the equipment, wherein the goal is related to a desired operating performance of the equipment; identifying, a plurality of sets of contiguous good reference segments, wherein a contiguous set of good reference segments comprises a plurality of operating time segments where the desired operating performance goal was achieved for a predetermined of time; identifying, a subset of sets comprising reference segments that are achievable from a current operating state of the equipment; selecting, a reference segment that is attainable based upon exogenous factors related to an operating environment of the equipment; and providing a recommendation to an operator of the

Abstract: A biometric system is tested to see whether a proposed use matches a configuration of the system. An enrolment input is received from an enrolling user, and compared with a system configuration model to obtain a configuration matching score value. The enrollment is then controlled based on a result of comparing the received enrollment input with the system configuration model. In the case of a voice biometric system, when a test input is received from a speaker, it is determined whether audio conditions applying to the test input correspond to system configuration conditions. Verification is performed by comparing the test input with a model of the speech of an enrolled user to generate a verification score for use in deciding whether to accept or reject the speaker, depending on whether it is determined that audio conditions applying to the test input correspond to the system configuration conditions.

Abstract: A method is described of identifying time-series signals that contain information useful for predicting impending event messages relating to one or more of safety, maintenance, and system operation information before they occur. The method includes loading a plurality of time-series signals with assigned signal name and associated time-series data into a machine-readable storage medium and grouping the plurality of time-series signals based on textual similarity of the corresponding signal names into a signal cluster.

Abstract: Systems and methods for efficiently detecting and coordinating step changes, trends, cycles, and bursts affecting lexical items within data streams are provided. Data streams can be sourced from documents that can optionally be labeled with metadata. Changes can be grouped across lexical and/or metavalue vocabularies to summarize the changes that are synchronous in time. The methods described herein can be applied either retrospectively to a corpus of data or in a streaming mode.

Abstract: In an embodiment, the disclosed technologies include automatically recognizing speech content of an audio stream that may contain multiple different classes of speech content, by receiving, by an audio capture device, an audio stream; outputting, by one or more classifiers, in response to an inputting to the one or more classifiers of digital data that has been extracted from the audio stream, score data; where a score of the score data indicates a likelihood that a particular time segment of the audio stream contains speech of a particular class; where the one or more classifiers use one or more machine-learned models that have been trained to recognize audio of one or more particular classes to determine the score data; using a sliding time window process, selecting particular scores from the score data; using the selected particular scores, determining and outputting one or more decisions as to whether one or more particular time segments of the audio stream contain speech of one or more particular classes

Abstract: A building management system includes connected equipment configured to measure a plurality of monitored variables and a predictive diagnostics system configured to receive the monitored variables from the connected equipment; generate a probability distribution of the plurality of monitored variables; determine a boundary for the probability distribution using a supervised machine learning technique to separate normal conditions from faulty conditions indicated by the plurality of monitored variables; separate the faulty conditions into sub-patterns using an unsupervised machine learning technique to generate a fault prediction model, each sub-pattern corresponding with a fault, and each fault associated with a fault diagnosis; receive a current set of the monitored variables from the connected equipment; determine whether the current set of monitored variables correspond with one of the sub-patterns of the fault prediction model to facilitate predicting whether a corresponding fault will occur; and determin

Abstract: The disclosure relates to a transmission device, comprising: a processor configured: to generate a multicarrier signal based on a combination of data symbols and reference symbols, wherein the multicarrier signal comprises a first plurality of inband subcarriers and a second plurality of out-of band (OOB) subcarriers, and to precode the multicarrier signal based on a mapping function with respect to the first plurality of inband subcarriers and the second plurality of out-of band subcarriers, wherein the mapping function is configured to mitigate the OOB subcarriers.

Abstract: A method executable via operation of configured processing circuitry may include constructing a mutual information graph for categorical data with respect to observed attributes of a plurality of entities described in terms of respective ones of the observed attributes by the categorical data, determining a clique tree correlating attributes having at least a threshold level of mutual dependence among the observed attributes, and determining a normality rating for an entity relative to the plurality of entities based on the clique tree.

Abstract: A user registration method and a device for a smart robot. The method comprises: conducting a voice dialogue with a new user to be registered, acquiring a user name of the user from the voice dialogue, and simultaneously collecting biological characteristic information that can uniquely identify the user; wherein the biological characteristic information comprises at least two different types of biological characteristic information, judging whether at least one type of the biological characteristic information satisfies a corresponding preset registration condition, and if yes, using the biological characteristic information that satisfies the preset registration condition as a characteristic template, establishing a correspondence relation between the characteristic template and the user name, and saving the correspondence relation, to complete the user registration.

Abstract: A speech synthesis dictionary delivery device that delivers a dictionary for performing speech synthesis to terminals, comprises a storage device for speech synthesis dictionary database that stores a first dictionary which includes an acoustic model of a speaker and is associated with identification information of the speaker, that stores a second dictionary which includes an acoustic model generated using voice data of a plurality of speakers, and that stores parameter sets of the speakers to be used with the second dictionary and which are associated with identification information of the speakers, a processor that determines one of the first dictionary and the second dictionary, which should be used in the terminal for a specified speaker, and an input output interface (I/F) that receives the identification information of a speaker transmitted from the terminal and then delivers at least one of a first dictionary, the second dictionary, and a parameter set of the second dictionary, on the basis of the rec

Abstract: Systems and methods are disclosed for generating internal state representations of a neural network during processing and using the internal state representations for classification or search. In some embodiments, the internal state representations are generated from the output activation functions of a subset of nodes of the neural network. The internal state representations may be used for classification by training a classification model using internal state representations and corresponding classifications. The internal state representations may be used for search, by producing a search feature from an search input and comparing the search feature with one or more feature representations to find the feature representation with the highest degree of similarity.

Abstract: To provide a system capable of appropriately proposing a search term candidate for each page of a document. Provided is a search document information storage device comprising: a vocabulary extraction means 3; a keyword storage means 5; a keyword extraction means 7; a topic term storage means 9; a topic term extraction means 11; a search term candidate extraction means 13; a search term candidate display means 17; a search term input means 19; and a document search information storage means 21.

Abstract: Methods, computer program products, and systems are presented. The methods include, for instance: obtaining the media file with a speech and identifying speakers on clusters separated by disfluencies and change of speakers. Clusters are re-segmented rearranged during diarization. Speaker identifications for the clusters in the media file is produced.

Abstract: Described is a system for classifying novel objects in imagery. In operation, the system extracts salient patches from a plurality of unannotated images using a multi-layer network. Activations of the multi-layer network are clustered into key attribute, with the key attributes being displayed to a user on a display, thereby prompting the user to annotate the key attributes with class label. An attribute database is then generated based on user prompted annotations of the key attributes. A test image can then be passed through the system, allowing the system to classify at least one object in the test image by identifying an object class in the attribute database. Finally, a device can be caused to operate or maneuver based on the classification of the at least one object in the test image.

Abstract: Embodiments of the present invention provide a system for automatically extracting conversational structure from a voice record based on lexical and acoustic features. The system also aggregates business-relevant statistics and entities from a collection of spoken conversations. The system may infer a coarse-level conversational structure based on fine-level activities identified from extracted acoustic features. The system improves significantly over previous systems by extracting structure based on lexical and acoustic features. This enables extracting conversational structure on a larger scale and finer level of detail than previous systems, and can feed an analytics and business intelligence platform, e.g. for customer service phone calls. During operation, the system obtains a voice record. The system then extracts a lexical feature using automatic speech recognition (ASR). The system extracts an acoustic feature.

Abstract: Methods, computer program products, and systems are presented. The methods include, for instance: obtaining the media file with a speech and identifying speakers on clusters separated by disfluencies and change of speakers. Clusters are re-segmented rearranged during diarization. Speaker identifications for the clusters in the media file is produced.

Abstract: The method comprises receive first audio comprising speech from a user of a computing device, detecting an end of speech in the first audio, generating an ASR result based, at least in part, on a portion of the first audio prior to the detected end of speech, determining whether a valid action can be performed by a speech-enabled application installed on the computing device using the ASR result, and processing second audio when it is determined that a valid action cannot be performed by the speech-enabled application using the ASR result.

Abstract: According to one embodiment, a voice keyword detection apparatus includes a memory and a circuit coupled with the memory. The circuit calculates a first score for a first sub-keyword and a second score for a second sub-keyword. The circuit detects the first and second sub-keywords based on the first and second scores. The circuit determines, when the first sub-keyword is detected from one or more first frames, to accept the first sub-keyword. The circuit determines, when the second sub-keyword is detected from one or more second frames, whether to accept the second sub-keyword based on a start time and/or an end time of the one or more first frames and a start time and/or an end time of the one or more second frames.

Abstract: According to an embodiment, a speech recognition result output device includes a storage and processing circuitry. The storage is configured to store a language model for speech recognition. The processing circuitry is coupled to the storage and configured to acquire a phonetic sequence, convert the phonetic sequence into a phonetic sequence feature vector, convert the phonetic sequence feature vector into graphemes using the language model, and output the graphemes.

Abstract: With reference to information storing a co-occurrence probability of each of plural words in association with each of distribution-destinations, the apparatus extracts, from a message to be distributed, an unknown-word that is not included in the plural words, where the co-occurrence probability indicates a probability that each word is included in a message to be distributed to each distribution-destination.

Abstract: A display system for an issue comprises an input unit, a display unit and an processing unit. The input unit receives an initial keyword corresponding to an issue. The display unit displays at least a derivative issue generated from the issue during a time period according to time-based characteristics. The processing unit coupled to the input unit and the display unit obtains tags of subject contents of web pages, and obtains a present keywords group according to co-occurrence correlation of the tags. The processing unit analyzes the correlation between the present keywords calculated based on social voice, analyzing overlap rate for the present keywords compared with the initial keywords, and compares correlation between the present keywords with correlation between the initial keywords calculated based on social voice, in order to determine whether at least one of the derivative issue is generated.

Abstract: A method includes accessing a first sensor log and a corresponding first reference log. Each of the first sensor log and the first reference log includes a series of measured values of a parameter according to a first time series. The method also includes accessing a second sensor log and a corresponding second reference log. Each of the second sensor log and the second reference log includes a series of measured values of a parameter according to a second time series. The method also includes dynamically time warping the first reference log and/or second reference log by a first transformation between the first time series and a common time-frame and/or a second transformation between the second time series and the common time-frame. The method also includes generating first and second warped sensor logs by applying the or each transformation to the corresponding ones of the first and second sensor logs.

Abstract: Systems and methods are disclosed for generating internal state representations of a neural network during processing and using the internal state representations for classification or search. In some embodiments, the internal state representations are generated from the output activation functions of a subset of nodes of the neural network. The internal state representations may be used for classification by training a classification model using internal state representations and corresponding classifications. The internal state representations may be used for search, by producing a search feature from an search input and comparing the search feature with one or more feature representations to find the feature representation with the highest degree of similarity.

Abstract: Systems, methods, devices, and other techniques for training and using a speaker verification neural network. A computing device may receive data that characterizes a first utterance. The computing device provides the data that characterizes the utterance to a speaker verification neural network. Subsequently, the computing device obtains, from the speaker verification neural network, a speaker representation that indicates speaking characteristics of a speaker of the first utterance. The computing device determines whether the first utterance is classified as an utterance of a registered user of the computing device. In response to determining that the first utterance is classified as an utterance of the registered user of the computing device, the device may perform an action for the registered user of the computing device.

Abstract: There is provided a system comprising a microphone, configured to receive an input speech from an individual, an analog-to-digital (A/D) converter to convert the input speech to digital form and generate a digitized speech, a memory storing an executable code and an age estimation database, a hardware processor executing the executable code to receive the digitized speech, identify a plurality of boundaries in the digitized speech delineating a plurality of phonemes in the digitized speech, extract a plurality of formant-based feature vectors from each phoneme in the digitized speech based on at least one of a formant position, a formant bandwidth, and a formant dispersion, compare the plurality of formant-based feature vectors with age determinant formant-based feature vectors of the age estimation database, determine the age of the individual when the comparison finds a match in the age estimation database, and communicate an age-appropriate response to the individual.

Abstract: A speech recognition method capable of automatic generation of phones according to the present invention includes: unsupervisedly learning a feature vector of speech data; generating a phone set by clustering acoustic features selected based on an unsupervised learning result; allocating a sequence of phones to the speech data on the basis of the generated phone set; and generating an acoustic model on the basis of the sequence of phones and the speech data to which the sequence of phones is allocated.

Abstract: A network device for implementing voice input comprises an input-obtaining module for obtaining voice input information, a sequence-determining module for determining an input character sequence corresponding to the voice input information based on a voice recognition model, an accuracy-determining module for determining appearance-probability information corresponding to word segments in the input character sequence so as to obtain accuracy information of the word segments, and a transmitting module for transmitting, to a user device, the input character sequence and the accuracy information of the word segments corresponding to the voice input information.

Abstract: Embodiments of the invention include method, systems and computer program products for role modeling. Aspects of the invention include receiving, by a processor, audio data, wherein the audio data includes a plurality of audio conversation for one or more speakers. The one or more segments for each of the plurality of audio conversations are partitioned. A speaker is associated with each of the one or more segments. The one or more segments for each of the plurality of audio conversations are labeled with roles utilizing a speaker recognition engine. Speakers are clustered based at least in part on a number of times the speakers are present in an audio conversation.

Abstract: Methods, systems, and apparatuses for audio event detection, where the determination of a type of sound data is made at the cluster level rather than at the frame level. The techniques provided are thus more robust to the local behavior of features of an audio signal or audio recording. The audio event detection is performed by using Gaussian mixture models (GMMs) to classify each cluster or by extracting an i-vector from each cluster. Each cluster may be classified based on an i-vector classification using a support vector machine or probabilistic linear discriminant analysis. The audio event detection significantly reduces potential smoothing error and avoids any dependency on accurate window-size tuning. Segmentation may be performed using a generalized likelihood ratio and a Bayesian information criterion, and the segments may be clustered using hierarchical agglomerative clustering. Audio frames may be clustered using K-means and GMMs.

Abstract: Speech recognition systems that use voice templates may create (or update) voice templates for a particular user by training (or re-training). If a training results in a vocabulary with similar voice templates, then the speech recognition system's performance may suffer. The present invention provides embraces methods for training a speech recognition system to prevent voice template similarity. In these methods, a trained word's voice template may be evaluated for similarity to other vocabulary templates prior to enrolling the voice template into the vocabulary. If template similarity is found, then a user may be prompted to retrain the system using an alternate word. Alternatively, the user may be prompted to retrain the system with the word spoken more clearly. This dynamic enrollment training analysis insures that all templates in the vocabulary are distinct.

Abstract: In a method of diarization of audio data, audio data is segmented into a plurality of utterances. Each utterance is represented as an utterance model representative of a plurality of feature vectors. The utterance models are clustered. A plurality of speaker models are constructed from the clustered utterance models. A hidden Markov model is constructed of the plurality of speaker models. A sequence of identified speaker models is decoded.

Abstract: Software that trains an artificial neural network for generating vector representations for natural language text, by performing the following steps: (i) receiving, by one or more processors, a set of natural language text; (ii) generating, by one or more processors, a set of first metadata for the set of natural language text, where the first metadata is generated using supervised learning method(s); (iii) generating, by one or more processors, a set of second metadata for the set of natural language text, where the second metadata is generated using unsupervised learning method(s); and (iv) training, by one or more processors, an artificial neural network adapted to generate vector representations for natural language text, where the training is based, at least in part, on the received natural language text, the generated set of first metadata, and the generated set of second metadata.

Abstract: Disclosed herein are methods of diarizing audio data using first-pass blind diarization and second-pass blind diarization that generate speaker statistical models, wherein the first pass-blind diarization is on a per-frame basis and the second pass-blind diarization is on a per-word basis, and methods of creating acoustic signatures for a common speaker based only on the statistical models of the speakers in each audio session.

Abstract: An audio apparatus includes a receiver configured to receive audio data and audio transducer position data for a plurality of audio transducers; and a renderer configured to render the audio data by generating audio transducer drive signals for the audio transducers from the audio data. Further, a clusterer is configured to cluster the audio transducers into a set of clusters in response to the audio transducer position data and to distances between audio transducers in accordance with a distance metric. A render controller is configured to adapt the rendering in response to the clustering. The apparatus is configured to select array processing techniques for specific subsets that contain audio transducers that are sufficiently close and allow automatic adaptation to audio transducer configurations thereby, e.g., allowing a user increased flexibility in positioning loudspeakers.

Abstract: A method for verifying at least one sound sample to be used in generating a sound detection model in an electronic device includes receiving a first sound sample; extracting a first acoustic feature from the first sound sample; receiving a second sound sample; extracting a second acoustic feature from the second sound sample; and determining whether the second acoustic feature is similar to the first acoustic feature.

Abstract: One embodiment of the present invention sets forth a technique for determining a set of sound parameters associated with a sound type. The technique includes receiving, via a network and from each a first plurality of remote computing devices, an audio recording of a first sound type and a descriptor associated with the first sound type. The technique further includes processing the audio recordings via a processor to determine a first set of sound parameters associated with the first sound type. The technique further includes receiving a request associated with the descriptor from at least one remote computing device and, in response, transmitting the first set of sound parameters associated with the first sound type to the at least one remote computing device.

Abstract: A statistical basic classification model of acoustic features generated for at least one reference object is automatically adapted by a data processing unit based on acoustic features of a noise generated by an object to be investigated to obtain an individually adapted statistical classification model. The data processing unit then classifies the state of the noise-generating object based on the individually adapted statistical classification model.

Abstract: Systems and methods to perform speaker clustering determine which audio segments appear to include sound generated by the same speaker. Speaker clustering is based on creating a graph in which a node represents an audio segment and an edge between two nodes represents a relationship and/or correspondence that reflects a probability, likelihood, or other indication that the two nodes represent audio segments of the same speaker. This graph is analyzed to detect individual communities of nodes that associate to an individual speaker.