Abstract: An aspect of the present disclosure relate to translating natural language queries into native database queries. For example, a database application receives a natural language query from a user, divides the query into clauses, determines keywords from the clauses, and maps keywords in the query to native query that can query a database. In the case that the natural language query includes keywords that are not mapped, the application resolves an ambiguity by communicating with a user device. In the case that a keyword cannot be mapped, the application can map the keyword to a custom index that includes synonyms derived using pair-wise generalization or other techniques. Once all keywords are mapped, the database application generates a native database query and obtains results from a database.

Abstract: The invention provides methods and devices for stereo encoding and decoding using complex prediction in the frequency domain. In one embodiment, a decoding method, for obtaining an output stereo signal from an input stereo signal encoded by complex prediction coding and comprising first frequency-domain representations of two input channels, comprises the upmixing steps of: (i) computing a second frequency-domain representation of a first input channel; and (ii) computing an output channel on the basis of the first and second frequency-domain representations of the first input channel, the first frequency-domain representation of the second input channel and a complex prediction coefficient. The upmixing can be suspended responsive to control data.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for audio classifiers. In one aspect, a method includes obtaining a plurality of video frames from a plurality of videos, wherein each of the plurality of video frames is associated with one or more image labels of a plurality of image labels determined based on image recognition; obtaining a plurality of audio segments corresponding to the plurality of video frames, wherein each audio segment has a specified duration relative to the corresponding video frame; and generating an audio classifier trained using the plurality of audio segment and the associated image labels as input, wherein the audio classifier is trained such that the one or more groups of audio segments are determined to be associated with respective one or more audio labels.

Abstract: Speech representing a desired video site or video subject is detected and digitized at a TV remote, and then sent to a TV. The TV or in some embodiments an Internet server communicating with the TV use speech recognition principles to recognize the speech, enter a database using the recognized speech as entering argument, and return a link to an Internet site hosting the desired video. The link can be displayed on the TV for selection thereof by a user to retrieve the video.

Abstract: Speech recognition may be improved by generating and using a topic specific language model. A topic specific language model may be created by performing an initial pass on an audio signal using a generic or basis language model. A speech recognition device may then determine topics relating to the audio signal based on the words identified in the initial pass and retrieve a corpus of text relating to those topics. Using the retrieved corpus of text, the speech recognition device may create a topic specific language model. In one example, the speech recognition device may adapt or otherwise modify the generic language model based on the retrieved corpus of text.

Abstract: Systems and methods are disclosed herein for improving audio conferencing services. One aspect relates to processing audio content of a conference. A first audio signal is received from a first conference participant, and a start and an end of a first utterance by the first conference participant are detected from the first audio signal. A second audio signal is received from a second conference participant, and a start and an end of a second utterance by the second conference participant is detected from the second audio signal. The second conference participant is provided with at least a portion of the first utterance, wherein at least one of start time, start point, and duration is determined based at least in part on the start, end, or both, of the second utterance.

Abstract: A system, apparatus, and method for processing a natural language, and a computer readable recording medium are provided. The system includes a user device configured to receive a compound or complex sentence and a natural language processing apparatus configured to generate a plurality of control commands for the user device based on whether operations intended by a user are sequentially performable in the compound or complex sentence received from the user device.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for speech recognition. One method includes obtaining an input acoustic sequence, the input acoustic sequence representing an utterance, and the input acoustic sequence comprising a respective acoustic feature representation at each of a first number of time steps; processing the input acoustic sequence using a first neural network to convert the input acoustic sequence into an alternative representation for the input acoustic sequence; processing the alternative representation for the input acoustic sequence using an attention-based Recurrent Neural Network (RNN) to generate, for each position in an output sequence order, a set of substring scores that includes a respective substring score for each substring in a set of substrings; and generating a sequence of substrings that represent a transcription of the utterance.

Abstract: Embodiments of the present invention perform speaker identification and verification by first prompting a user to speak a phrase that includes a common phrase component and a personal identifier. Then, the embodiments decompose the spoken phrase to locate the personal identifier. Finally, the embodiments identify and verify the user based on the results of the decomposing.

Abstract: The subject matter of this specification can be embodied in, among other things, a method that includes receiving geographical information derived from a non-verbal user action associated with a first computing device. The non-verbal user action implies an interest of a user in a geographic location. The method also includes identifying a grammar associated with the geographic location using the derived geographical information and outputting a grammar indicator for use in selecting the identified grammar for voice recognition processing of vocal input from the user.

Abstract: Provided are techniques for avoiding sentiment model overfitting in a machine language model. A current list of keywords in a current sentiment model can be updated to create a proposed list of keywords in a proposed sentiment model. Machine-generated sentiment results, based on the proposed sentiment model, are presented to identify model overfitting, without revising the current set of keywords. The proposed set of keywords can be edited, and when overfitting is not present, the current list of keywords is replaced by the proposed list of keywords.

Abstract: The present invention is provided with: a voice input section that receives a remark (a question) via a voice signal; a reply creation section that creates a voice sequence of a reply (response) to the remark; a pitch analysis section that analyzes the pitch of a first segment (e.g., word ending) of the remark; and a voice generation section (a voice synthesis section, etc.) that generates a reply, in the form of voice, represented by the voice sequence. The voice generation section controls the pitch of the entire reply in such a manner that the pitch of a second segment (e.g., word ending) of the reply assumes a predetermined pitch (e.g., five degrees down) with respect to the pitch of the first segment of the remark. Such arrangements can realize synthesis of replying voice capable of giving a natural feel to the user.

Abstract: The present disclosure is related to devices, systems, and methods for taking an action in response to detection of an unsupported language in a log. An example system can include instructions to analyze a log of a log source, detect an unsupported language in the log, and take an action in response to detecting the unsupported language in the log.

Abstract: A method and a device for providing information are provided. The method includes obtaining, via a communication terminal, speech information provided by at least one participant in a communication conducted using the communication terminal via the communication terminal; obtaining state information indicating a current state of the at least one participant by using at least one from among the speech information and data stored in advance; determining a virtual speech to be provided to the at least one participant based on the speech information and the state information; and outputting the determined virtual speech.

Abstract: In a pulse encoding and decoding method and a pulse codec, more than two tracks are jointly encoded, so that free codebook space in the situation of single track encoding can be combined during joint encoding to become code bits that may be saved. Furthermore, a pulse that is on each track and required to be encoded is combined according to positions, and the number of positions having pulses, distribution of the positions that have pulses on the track, and the number of pulses on each position that has a pulse are encoded separately, so as to avoid separate encoding performed on multiple pulses of a same position, thereby further saving code bits.

Abstract: Methods and systems for training a conversation-classification model are disclosed. A first set of conversations in a source domain and a second set of conversation in a target domain are received. Each of the first set of conversations has an associated predetermined tag. One or more features are extracted from the first set of conversations and from the second set of conversations. Based on the similarity of content in the first set of conversations and the second set of conversations, a first weight is assigned to each conversation of the first set of conversations. Further, a second weight is assigned to the one or more features of the first set of conversations based on the similarity of the one or more features of the first set of conversations and of the second set of conversations. A conversation-classification model is trained based on the first weight and the second weight.

Abstract: A system for training a model to predict a sequence (e.g. a sequence of words) given a context is disclosed. A model can be trained to make these predictions using a combination of individual predictions compared to base truth and sequences of predictions based on previous predictions, where the resulting sequence is compared to the base truth sequence. In particular, the model can initially use the individual predictions to train the model. The model can then be further trained over the training data in multiple iterations, where each iteration includes two processes for each training element. In the first process, an initial part of the sequence is predicted, and the model and model parameters are updated after each prediction. In the second process, the entire remaining amount of the sequence is predicted and compared to the corresponding training sequence to adjust model parameters to encourage or discourage each prediction.

Abstract: The present invention is directed to a deep neural network (DNN) having a triplet network architecture, which is suitable to perform speaker recognition. In particular, the DNN includes three feed-forward neural networks, which are trained according to a batch process utilizing a cohort set of negative training samples. After each batch of training samples is processed, the DNN may be trained according to a loss function, e.g., utilizing a cosine measure of similarity between respective samples, along with positive and negative margins, to provide a robust representation of voiceprints.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for audio classifiers. In one aspect, a method includes obtaining a plurality of video frames from a plurality of videos, wherein each of the plurality of video frames is associated with one or more image labels of a plurality of image labels determined based on image recognition; obtaining a plurality of audio segments corresponding to the plurality of video frames, wherein each audio segment has a specified duration relative to the corresponding video frame; and generating an audio classifier trained using the plurality of audio segment and the associated image labels as input, wherein the audio classifier is trained such that the one or more groups of audio segments are determined to be associated with respective one or more audio labels.

Abstract: A system configured to process speech commands may classify incoming audio as desired speech, undesired speech, or non-speech. Desired speech is speech that is from a same speaker as reference speech. The reference speech may be obtained from a configuration session or from a first portion of input speech that includes a wakeword. The reference speech may be encoded using a recurrent neural network (RNN) encoder to create a reference feature vector. The reference feature vector and incoming audio data may be processed by a trained neural network classifier to label the incoming audio data (for example, frame-by-frame) as to whether each frame is spoken by the same speaker as the reference speech. The labels may be passed to an automatic speech recognition (ASR) component which may allow the ASR component to focus its processing on the desired speech.