Abstract: A system and method enable one to set a target duration of a desired synthesized utterance without removing or adding spoken content. Without changing the spoken text, the voice characteristics may be kept the same or substantially the same. Silence adjustment and interpolation may be used to alter the duration while preserving speech characteristics. Speech may be translated prior to a vocoder step, pursuant to which the translated speech is constrained by the original audio duration, while mimicking the speech characteristics of the original speech.

Abstract: Embodiments of the present invention relate to computer-implemented methods, systems, and computer program products for managing a conversational system. In one embodiment, a computer-implemented method comprises: obtaining, by a device operatively coupled to one or more processors, a first message sequence comprising messages involved in a conversation between a user and a conversation server; obtaining, by the device, a conversation graph indicating an association relationship between messages involved in a conversation; and in response to determining that the first message sequence is not matched in the conversation graph, updating, by the device, the conversation graph with a second message sequence, the second message sequence being generated based on a knowledge library including expert knowledge that is associated with a topic of the conversation.

Abstract: Methods and systems for generating voices based on characteristics of an avatar. One or more characteristics of an avatar are obtained and one or more parameters of a voice synthesizer for generating a voice corresponding to the one or more avatar characteristics are determined. The voice synthesizer is configured based on the one or more parameters and a voice is generated using the parameterized voice synthesizer.

Abstract: The technology disclosed herein summarizes a document using a dictionary derived from tokens within the document itself. In a particular implementation, a method provides identifying a first document for summarization and inputting the first document into a natural language model. The natural language model is configured to summarize the first document using words from a first dictionary compiled based on tokens from the first document. The method further provides receiving a first summary output by the natural language model after the natural language model summarizes the first document.

Abstract: The disclosed technology relates to a process for automatically training a machine learning algorithm to recognize a custom wake word. By using different text-to-speech services, input providing a custom wake word to a text to speech service can be used in order to generate different speech samples covering different variations in how the custom wake word can be pronounced. These samples are automatically generated and are subsequently used to train the wake word detection algorithm that will be used by the computing device to recognize and detect when the custom wake word is uttered by any user nearby a computing device for the purposes of initiating a virtual assistant. In a further embodiment, “white-listed” words (e.g different words that are pronounced similar to the custom wake word) are also identified and trained in order to minimize the occurrence of erroneously initiating the virtual assistant.

Abstract: Embodiments of a computer-implemented system for improving classification of data associated with the deep web or dark net are disclosed.

Abstract: A voice interaction system performs a voice interaction with a user. The voice interaction system includes: topic detection means for estimating a topic of the voice interaction and detecting a change in the topic that has been estimated; and ask-again detection means for detecting, when the change in the topic has been detected by the topic detection means, the user's voice as ask-again by the user based on prosodic information on the user's voice.

Abstract: A computing device receives a natural language input from a user. The computing device routes the natural language input from an active domain node of multiple domain nodes of a multi-domain context-based hierarchy to a leaf node of the domain nodes by selecting a parent domain node in the hierarchy until an off-topic classifier labels the natural language input as in-domain and then selecting a subdomain node in the hierarchy until an in-domain classifier labels the natural language input with a classification label, each of the plurality of domain nodes comprising a respective off-topic classifier and a respective in-domain classifier trained for a respective domain node. The computing device outputs the classification label determined by the leaf node.

Abstract: A technique for data augmentation for speech data is disclosed. Original speech data including a sequence of feature frames is obtained. A partially prolonged copy of the original speech data is generated by inserting one or more new frames into the sequence of the feature frames. The partially prolonged copy is output as augmented speech data for training an acoustic model for training an acoustic model.

Abstract: A method and an apparatus for extracting information are provided. The method according to an embodiment includes: receiving and parsing voice information of a user to generate text information corresponding to the voice information; extracting to-be-recognized contact information from the text information; acquiring an address book of the user, the address book including at least two pieces of contact information; generating at least two types of matching information based on the to-be-recognized contact information; determining, for each of the at least two types of matching information, a matching degree between the to-be-recognized contact information and each of at least two pieces of contact information based on the type of matching information; and extracting contact information matching the to-be-recognized contact information from the address book based on the determined matching degree.

Abstract: An electronic apparatus is provided. The electronic apparatus includes an inputter comprising input circuitry, a voice receiver comprising voice receiving circuitry, a storage, and a processor configured to: provide a guide prompting a user utterance based on user authentication being performed according to user information input through the inputter, generate a speaker recognition model corresponding to the user information based on a voice corresponding to the guide being received through the voice receiver, store the speaker recognition model in the storage, and identify a user corresponding to a voice received through the voice receiver based on the speaker recognition model updated by comparing a voice received through the voice receiver with the speaker recognition model.

Abstract: In some implementations, a computing system is provided. The computing system includes a device. The device includes a non-volatile memory divided into a plurality of memory sub-arrays. Each memory sub-array comprises a plurality of selectable locations. A plurality of data processing units are communicatively coupled to the non-volatile memory in the absence of a central processing unit of the computing system. The data processing unit is assigned to process data of a memory sub-array, and configured to store the first data object in the non-volatile memory receive a first data object via a communication interface. The first data object comprises a first content and is associated with a first set of keywords. The data processing unit is also configured to add the first set of keywords to a local dictionary. The local dictionary is stored in the non-volatile memory. The data processing unit is further configured to determine whether the first data object is related to one or more other data objects.

Abstract: An electronic device includes an audio capture device receiving audio input. The electronic device includes one or more processors, operable with the audio capture device, and configured to execute a control operation in response to a device command preceded by a trigger phrase identified in the audio input when in a first mode of operation. The one or more processors transition from the first mode of operation to a second mode of operation in response to detecting a predefined operating condition of the electronic device. In the second mode of operation, the one or more processors execute the control operation without requiring the trigger phrase to precede the device command.

Abstract: An attribute identification technology that can reject an attribute identification result if the reliability thereof is low is provided. An attribute identification device includes: a posteriori probability calculation unit 110 that calculates, from input speech, a posteriori probability sequence {q(c, i)} which is a sequence of the posteriori probabilities q(c, i) that a frame i of the input speech is a class c; a reliability calculation unit 120 that calculates, from the posteriori probability sequence {q(c, i)}, reliability r(c) indicating the extent to which the class c is a correct attribute identification result; and an attribute identification result generating unit 130 that generates an attribute identification result L of the input speech from the posteriori probability sequence {q(c, i)} and the reliability r(c).

Abstract: Systems and methods are described herein for disambiguating a voice search query that contains a command keyword by determining whether the user spoke a quotation from a content item and whether the user mimicked or approximated the way the quotation is spoken in the content item. The voice search query is transcribed into a string, and an audio signature of the voice search query is identified. Metadata of a quotation matching the string is retrieved from a database that includes audio signature information for the string as spoken within the content item. The audio signature of the voice search query is compared with the audio signature information in the metadata to determine whether the audio signature matches the audio signature information in the quotation metadata. If a match is detected, then a search result comprising an identifier of the content item from which the quotation comes is generated.

Abstract: Systems and methods are provided for detecting inaccuracy in a product title, comprising identifying, by running a string algorithm on a title associated with a product, at least one product type associated with the product, predicting, using a machine learning algorithm, at least one product type associated with the product based on the title, detecting an inaccuracy in the title, based on at least one of the identification or the prediction, and outputting, to a remote device, a message indicating that the title comprises the inaccuracy. Running the string algorithm may comprise receiving a set of strings, generating a tree based on the received set of strings, receiving the title, and traversing the generated tree using the title to find a match. Using the machine learning algorithm may comprise identifying words in the title, learning a vector representation for each character n-gram of each word, and summing each character n-gram.

Abstract: A method of processing audio communications over a network, comprising: at a first client device: receiving a first audio transmission from a second client device that is provided in a source language distinct from a default language associated with the first client device; obtaining current user language attributes for the first client device that are indicative of a current language used for the communication session at the first client device; if the current user language attributes suggest a target language currently used for the communication session at the first client device is distinct from the default language associated with the first client device: obtaining a translation of the first audio transmission from the source language into the target language; and presenting the translation of the first audio transmission in the target language to a user at the first client device.

Abstract: A method for learning a task includes capturing first information associated with at least one application executed by an electronic device. A sequence of user interface interactions for the at least one application is recorded. Second information are extracted from the sequence of user interface interactions. Events, action or a combination thereof are filtered from the second information using the first information. Recognition is performed on each element from the first information to generate a semantic ontology. An executable sequential event task bytecode is generated from each element of the semantic ontology and the filtered second information.

Abstract: A method of communication includes determining, at a mobile device, a speech quality metric for an incoming speech signal associated with a voice call. The speech quality metric is based on an environment of the mobile device. The method also includes converting incoming speech associated with the incoming speech signal to text in response to a determination that the speech quality metric fails to satisfy a speech quality metric threshold. The method further includes displaying the text at a display screen of the mobile device during the voice call.

Abstract: A log spectral envelope sequence L0, L1, . . . , LN?1 and an envelope code for the log spectral envelope sequence L0, L1, . . . , LN?1 are obtained. The log spectral envelope sequence L0, L1, . . . , LN?1 is an integer value sequence corresponding to binary logarithms of respective sample values of a spectral envelope sequence and is an integer value sequence whose total sum is 0. For a quantized spectral sequence {circumflex over (?)}X0, {circumflex over (?)}X1, . . . , {circumflex over (?)}XN?1, a smoothed spectral sequence ˜X0, ˜X1, . . . , ˜XN?1 is obtained by: for {circumflex over (?)}Xk with Lk being a positive value, adopting {circumflex over (?)}Xk with Lk digits from its least significant digit removed as ˜Xk; for {circumflex over (?)}Xk with Lk being a negative value, adopting {circumflex over (?)}Xk with ?Lk digits added to its least significant digit in accordance with a predefined rule as ˜Xk; and when Lk is 0, adopting {circumflex over (?)}Xk as ˜Xk.