Abstract: An explainable artificially intelligent (XAI) application contains an ordered sequence of artificially intelligent software modules. When an input dataset is submitted to the application, each module generates an output dataset and an explanation that represents, as a set of Boolean expressions, reasoning by which each output element was chosen. If any pair of explanations are determined to be semantically inconsistent, and if this determination is confirmed by further determining that an apparent inconsistency was not a correct response to an unexpected characteristic of the input dataset, nonzero inconsistency scores are assigned to inconsistent elements of the pair of explanations.

Abstract: A device for changing a speech recognition sensitivity for speech recognition can include a memory and a processor configured to obtain a first plurality of speech data input at different times, apply a pre-trained speech recognition model to the first plurality of speech data at a plurality of different speech recognition sensitivities, obtain a first speech recognition sensitivity from among the plurality of different speech recognition sensitivities based on the pre-trained speech recognition model and the plurality of different speech recognition sensitivities, the first speech recognition sensitivity corresponding to an optimal speech recognition sensitivity at which a speech recognition success rate of the speech recognition model satisfies a set first recognition success rate criterion, and change a setting of the speech recognition sensitivity based on the first speech recognition sensitivity obtained from among the plurality of different speech recognition sensitivities.

Abstract: In one embodiment, a method includes receiving, from a client system associated with a first user, a first audio input. The method includes generating multiple transcriptions corresponding to the first audio input based on multiple automatic speech recognition (ASR) engines. Each ASR engine is associated with a respective domain out of multiple domains. The method includes determining, for each transcription, a combination of one or more intents and one or more slots to be associated with the transcription. The method includes selecting, by a meta-speech engine, one or more combinations of intents and slots from the multiple combinations to be associated with the first user input. The method includes generating a response to the first audio input based on the selected combinations and sending, to the client system, instructions for presenting the response to the first audio input.

Abstract: Techniques for processing complex natural language inputs are described. A complex natural language input may be semantically tagged and parsed to identify individual clauses in the complex natural language input. An execution graph may be generated to represent the clauses and their dependencies. Nodes of the execution graph may be processed using NLU processing and/or a knowledge graph or other information storage and retrieval techniques, and results of such processing may be used to update clause variables with specific entities in the execution graph.

Abstract: Provided is an artificial intelligence (AI) device for recognizing speech of user. The AI apparatus includes: a microphone; and a processor configured to: receive, via the microphone, a sound signal corresponding to speech of the user, recognize the speech from the sound signal using a language model, determine an intention of the user based on the recognition result, determine whether the determination of the intention is successful, obtain a user's application usage log if the determination of the intention is not successful, and update the language model using the obtained user's application usage log.

Abstract: A system is provided for reducing friction during user interactions with a natural language processing system, such as voice assistant systems. The system determines a pre-trained model using dialog session data corresponding to multiple user profiles. The system determines a fine-tuned model using the pre-trained model and a fine-tuning dataset that corresponds to a particular task, such as query rewriting. The system uses the fine-tuned model to process a user input and determine an alternative representation of the input that can result in a desired response from the natural language processing system.

Abstract: Techniques are described herein for enabling the use of “dynamic” or “context-specific” hot words for an automated assistant. In various implementations, an automated assistant may be operated at least in part on a computing device. Audio data captured by a microphone may be monitored for default hot word(s). Detection of one or more of the default hot words may trigger transition of the automated assistant from a limited hot word listening state into a speech recognition state. Transition of the computing device into a given state may be detected, and in response, the audio data captured by the microphone may be monitored for context-specific hot word(s), in addition to or instead of the default hot word(s). Detection of the context-specific hot word(s) may trigger the automated assistant to perform a responsive action associated with the given state, without requiring detection of default hot word(s).

Abstract: A system and method are presented for adaptive skill level assignments of agents in contact center environments. A client and a service collaborate to automatically determine the effectiveness of an agent handling an interaction that has been routed using skills-based routing. Evaluation operations may be performed including emotion detection, transcription of audio to text, keyword analysis, and sentiment analysis. The results of the evaluation are aggregated with other information such as the interaction's duration, agent skills and agent skill levels, and call requirement skills and skill levels, to update the agent's profile which is then used for subsequent routing operations.

Abstract: In one embodiment, a method includes receiving, from a client system associated with a first user, a first audio input. The method includes generating multiple transcriptions corresponding to the first audio input based on multiple automatic speech recognition (ASR) engines. Each ASR engine is associated with a respective domain out of multiple domains. The method includes determining, for each transcription, a combination of one or more intents and one or more slots to be associated with the transcription. The method includes selecting, by a meta-speech engine, one or more combinations of intents and slots from the multiple combinations to be associated with the first user input. The method includes generating a response to the first audio input based on the selected combinations and sending, to the client system, instructions for presenting the response to the first audio input.

Abstract: An automated training method enhances conversational effectiveness of a system participant interacting with a virtual client. The method operates on received speech input to produce text. A speech effectiveness analysis processor receives both the speech and text and produces quantified metrics measuring speech effectiveness based on the input parameters. Personalized feedback is generated based on the quantified metrics and is transmitted to the system participant. A virtual client dialog is generated using a dynamic simulation processor. The virtual client dialog is based on the quantified metrics, such that the virtual client responds to the speech input using the generated dialog.

Abstract: There is provided an information processing device and an information processing method that enable the intention of a speech of a user to be estimated more accurately. The information processing device includes: a detection unit configured to detect a breakpoint of a speech of a user on the basis of a result of recognition that is to be obtained during the speech of the user; and an estimation unit configured to estimate an intention of the speech of the user on the basis of a result of semantic analysis of a divided speech sentence obtained by dividing a speech sentence at the detected breakpoint of the speech. The present technology can be applied, for example, to a speech dialogue system.

Abstract: Provided are techniques for domain specific correction of output from automatic speech recognition. An output of an automatic speech recognition engine is received. An alphanumeric sequence is extracted from the output, where the alphanumeric sequence represents an erroneous translation by the automatic speech recognition engine. Candidates for the alphanumeric sequence are generated. The candidates are ranked based on scores associated with the candidates. A candidate of the candidates having a highest score of the scores is selected. The output is corrected by replacing the alphanumeric sequence with the selected candidate. The corrected output is returned.

Abstract: Embodiments of the present disclosure provide a method and an apparatus for identifying a key phrase in audio, a device and a computer readable storage medium. The method for identifying a key phrase in audio includes obtaining audio data to be identified. The method further includes identifying the key phrase in the audio data using a trained key phrase identification model. The key phrase identification model is trained based on first training data for identifying feature information of words in a first training text and second training data for identifying the key phrase in a second training text. In this way, embodiments of the present disclosure can accurately and efficiently identify key information in the audio data.

Abstract: Messages are processed to generate effectiveness predictions and/or other insights associated with the messages. Candidate messages are processed through a natural language processing (NLP) component to parse the candidate message into message elements for further processing. The message elements are converted to a vector or set of vectors, which are provided as input to a machine learning model to make predictions of message effectiveness. A contribution score can be made for each message element of the candidate message, which may be indicative of the importance or relevance for the individual message element to the overall predicted message effectiveness. Other message elements not originally within the message can be provided as candidates to replace message elements already located within the message. In this way, a message that is likely to be effective, such being likely to have a high conversion rate, can be published or otherwise distributed.

Abstract: A non-transitory computer-readable storage medium storing a program that causes a processor included in a computer mounted on a sound source direction estimation device to execute a process, the process includes calculating a sound pressure difference between a first voice data acquired from a first microphone and a second voice data acquired from a second microphone and estimating a sound source direction of the first voice data and the second voice data based on the sound pressure difference, outputting an instruction to execute a voice recognition on the first voice data or the second voice data in a language corresponding to the estimated sound source direction, and controlling a reference for estimating a sound source direction based on the sound pressure difference, based on a time length of the voice data used for the voice recognition based on the instruction and a voice recognition time length.

Abstract: In order to improve the accuracy of ASR, an utterance is transcribed using a plurality of language models, such as for example, an N-gram language model and a neural language model. The language models are trained separately. They each output a probability score or other figure of merit for a partial transcription hypothesis. Model scores are interpolated to determine a hybrid score. While recognizing an utterance, interpolation weights are chosen or updated dynamically, in the specific context of processing. The weights are based on dynamic variables associated with the utterance, the partial transcription hypothesis, or other aspects of context.

Abstract: A personalized text-to-speech system configured to perform speaker adaption is disclosed. The TTS system includes an acoustic model comprising a base neural network and a differential neural network. The base neural network is configured to generate acoustic parameters corresponding to a base speaker or voice actor, while the differential neural network is configured to generate acoustic parameters corresponding to differences between acoustic parameters of the base speaker and a particular target speaker. The output of the acoustic model is then a weighted linear combination of the output from the base neural network and differential neural network. The base neural network and differential neural network share a first input layer and first plurality of hidden layers. Thereafter, the base neural network further comprises a second plurality of hidden layers and output layer. In parallel, the differential neural network further comprises a third plurality of hidden layers and separate output layer.

Abstract: An artificial intelligence apparatus for recognizing speech by correcting misrecognized word includes a microphone and a processor. The processor is configured to obtain, via the microphone, speech data including speech of a user, convert the speech data into text by using an acoustic model and a language model, determine whether an uncertain recognition exists in an acoustic recognition result according to the acoustic model, determine whether the converted text is a normal sentence by using a natural language processing model if an uncertain recognition exists in the acoustic recognition result, determine a sentence most similar to the converted text among sentences pre-learned by using the language model if the converted text is not a normal sentence, replace the converted text with the determined most similar sentence, and generate a speech recognition result corresponding to the speech data by using the converted text.

Abstract: Techniques are described herein for enabling the use of “dynamic” or “context-specific” hot words to invoke an automated assistant. In various implementations, an automated assistant may be executed in a default listening state at least in part on a user's computing device(s). While in the default listening state, audio data captured by microphone(s) may be monitored for default hot words. Detection of the default hot word(s) transitions of the automated assistant into a speech recognition state. Sensor signal(s) generated by hardware sensor(s) integral with the computing device(s) may be detected and analyzed to determine an attribute of the user. Based on the analysis, the automated assistant may transition into an enhanced listening state in which the audio data may be monitored for enhanced hot word(s). Detection of enhanced hot word(s) triggers the automated assistant to perform a responsive action without requiring detection of default hot word(s).

Abstract: A method of training machine learning models (MLMs). An issue vector is generated using an issue MLM to generate a first output including first embedded natural language issue statements. An action vector is generated using an action MLM to generate a second output including related embedded natural language action statements. The issue and action MLMs are of a same type. An inner product of the first and second output is calculated, forming a third output. The third output is processed according to a sigmoid gate process to predict whether a given issue statement and corresponding action statement relate to a same call, resulting in a fourth output. A loss function is calculated from the fourth output by comparing the fourth output to a known result. The issue MLM and the action MLM are modified using the loss function to obtain a trained issue MLM and a trained action MLM.