Abstract: A voice detection method which makes it possible to detect the presence of voice signals in an noisy acoustic signal x(t) from a microphone, including the following consecutive steps: calculating a detection function FD(?) based on calculating a difference function D(?) varying in accordance with the shift ? on an integration window with length W starting at the time t0, with: a step of adapting the threshold in said current interval, in accordance with values calculated from the acoustic signal x(t) established in said current interval; searching for the minimum of the detection function FD(?) and comparing the minimum with a threshold, for (?) varying in a predetermined time interval referred to as current interval so as to detect the possible presence of a fundamental frequency F0 that is characteristic of a voice signal in said current interval.

Abstract: Translations for items of user-generated content are received from one or more users associated with a social networking system and/or one or more machine-generated translations. Each translation may translate a content item from a source or original language to a target language. Votes regarding the suitability of the translations are received from one or more users. In one aspect, the option to provide translations and/or vote on translations is provided to those users that are determined to be competent in the source language of the item of user-generated content. Determination of whether a user is competent in the source language can be based on the user profile of the user, the user-generated content of the user, the language competencies of the user's connections, and/or other social signals.

Abstract: One embodiment provides a method for generating a natural language resource using a parallel corpus, the method including: utilizing at least one processor to execute computer code that performs the steps of: receiving, from a parallel corpus, natural language text in a source language and a corresponding translation of the natural language text in a target language, wherein the natural language text in the source language comprises linguistic annotations; projecting the linguistic annotations from the source language natural language text to the target language natural language text; applying one or more filters to remove at least one projected linguistic annotation from the target language natural language text that results in at least one error; selecting at least one target language natural language text having substantially complete linguistic annotations; training a machine learning model using the selected at least one target language natural language text and annotations; and adding, using the trained

Abstract: An information processing device includes a first information processing unit, a communication unit, and a control unit. The first information processing unit performs predetermined information processing on input data to generate first processing result data. The communication unit is capable of receiving second processing result data generated by a second information processing unit capable of executing the same kind of information processing as the information processing on the input data under a condition with higher versatility. The control unit selects either the first processing result data or the second processing result data according to the use environment of the device.

Abstract: Embodiments of the present invention provide an approach for estimating the accuracy of a transcription of a voice recording. Specifically, in a typical embodiment, each word of a transcription of a voice recording is checked against a customer-specific dictionary and/or a common language dictionary. The number of words not found in either dictionary is determined. An accuracy number for the transcription is calculated from the number of said words not found and the total number of words in the transcription.

Abstract: An exemplary computer system configured to user multiple automatic speech recognizers (ASRs) with a plurality of language and acoustic models to increase the accuracy of speech recognition.

Abstract: Features are disclosed for modifying a statistical model to more accurately discriminate between classes of input data. A subspace of the total model parameter space can be learned such that individual points in the subspace, corresponding to the various classes, are discriminative with respect to the classes. The subspace can be learned using an iterative process whereby an initial subspace is used to generate data and maximize an objective function. The objective function can correspond to maximizing the posterior probability of the correct class for a given input. The initial subspace, data, and objective function can be used to generate a new subspace that better discriminates between classes. The process may be repeated as desired. A model modified using such a subspace can be used to classify input data.

Abstract: A text-reading device includes: a visual line direction detection device for a driver; a memory that stores the visual line direction when the driver looks at a display device; a gaze determination device that determines that the driver gazes the display device when a state that the detected visual line direction coincides with the stored visual line direction continues for predetermined time or longer; a voice conversion device that outputs text information of the display device as a voice signal based on an instruction; and a reading control device that inputs the instruction when the driver gazes the display device while the display device displays the text information, and the vehicle starts to move.

Abstract: An apparatus for enhancing a speech signal includes a spectral analysis circuit that generates time-frequency spectra of signals from a microphone array, a signal separation filter, and a hardware processor. The hardware processor identifies a key phrase in the signals and aligns a template of the key phrase to time-frequency spectra of a microphone in the array. A mask is generated from the aligned template and applied to the time-frequency spectra. The masked spectra are used to determine acoustic parameters that, in turn, are used to generate coefficients for configuring the source separation filter which is then used to process the spectra to generate an enhanced speech signal. The apparatus maybe used as a front end for a speech recognition engine, for example.

Abstract: In this invention, the design of the Huffman table can be done offline with a large input sequence database. The range of the quantization indices (or differential indices) for Huffman coding is identified. For each value of range, all the input signal which have the same range will be gathered and the probability distribution of each value of the quantization indices (or differential indices) within the range is calculated. For each value of range, one Huffman table is designed according to the probability. And in order to improve the bits efficiency of the Huffman coding, apparatus and methods to reduce the range of the quantization indices (or differential indices) are also introduced.

Abstract: A method of recognizing continuous digits uttered by a speaker using an automatic speech recognition (ASR) system includes receiving continuous digits via a microphone as speech from a user; detecting that recognition of one or more of the continuous digits falls below a predetermined confidence threshold; prompting the user to identify the continuous digits using a body gesture; detecting the body gesture made by the user; and identifying one or more of the continuous digits based on the body gesture.

Abstract: A method of presenting translated content items is disclosed. It is detected that a content item has been captured by a device of a user. It is identified that the content item is a candidate content item for translation. The candidate content item is translated; and the translated candidate content item is presented via a user interface of a wearable display of the device.

Abstract: The present disclosure relates to an electronic device and a method for operating same, which can divide a recognition section for voice recognition, process different types of voice recognition corresponding to the divided recognition sections, and improve a voice recognition rate. According to an embodiment, a method for recognizing a voice by an electronic device may include: detecting a voice input; determining the direction of the voice and a beamforming direction. Voice recognition is based on the voice when the direction of the voice and the beamforming direction correspond to each other. Different types of voice recognition may be performed for voices input in a first section and a second section, based on divided types of voice input sections, when the direction of the voice and the beamforming direction do not correspond to each other.

Abstract: One or more embodiments present positional information associated with a text to a user. In one embodiment, a determination is made that at least one line from a digital representation of text has been selected. Another determination is made that the line is associated with a set of positional information. The set of positional information is presented on a digital representation of a venue along with the presentation of the line of text.

Abstract: Embodiments include methods and systems for improving an acoustic model. Aspects include acquiring a first standard deviation value by calculating standard deviation of a feature from first training data and acquiring a second standard deviation value by calculating standard deviation of a feature from second training data acquired in a different environment from an environment of the first training data. Aspects also include creating a feature adapted to an environment where the first training data is recorded, by multiplying the feature acquired from the second training data by a ratio obtained by dividing the first standard deviation value by the second standard deviation value. Aspects further include reconstructing an acoustic model constructed using training data acquired in the same environment as the environment of the first training data using the feature adapted to the environment where the first training data is recorded.

Abstract: Embodiments include methods and systems for improving an acoustic model. Aspects include acquiring a first standard deviation value by calculating standard deviation of a feature from first training data and acquiring a second standard deviation value by calculating standard deviation of a feature from second training data acquired in a different environment from an environment of the first training data. Aspects also include creating a feature adapted to an environment where the first training data is recorded, by multiplying the feature acquired from the second training data by a ratio obtained by dividing the first standard deviation value by the second standard deviation value. Aspects further include reconstructing an acoustic model constructed using training data acquired in the same environment as the environment of the first training data using the feature adapted to the environment where the first training data is recorded.

Abstract: A frequency domain converter is configured to create a plurality of pieces of frequency domain information by individually converting a plurality of input audio signals, which is acquired at different positions, into frequency domain information. A relative value calculator is configured to calculate a relative value of time frequency components of at least one set of frequency domain information among the plurality of pieces of frequency domain information. A signal determiner is configured to determine whether or not each of the input audio signals includes an audio signal component, which is emitted from a predetermined position, based on whether or not the relative value is included in a range specified and based on a relative threshold value stored in a memory in advance.

Abstract: A processing unit of a voice processing apparatus first generates a target voice signal in a time domain by adjusting a fundamental frequency of a target voice signal to a fundamental frequency of an initial voice signal, so as to generate a spectrum of the target voice signal after pitch is adjusted. Second, the processing unit reallocates, along a frequency axis, the spectrum of the target voice characteristics by having the spectrum correspond to each of the fundamental frequencies of the initial voice signal.

Abstract: A method includes providing a deep neural network acoustic model, receiving audio data including one or more utterances of a speaker, extracting a plurality of speech recognition features from the one or more utterances of the speaker, creating a speaker identity vector for the speaker based on the extracted speech recognition features, and adapting the deep neural network acoustic model for automatic speech recognition using the extracted speech recognition features and the speaker identity vector.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for enhancing speech recognition accuracy. In one aspect, a method includes receiving voice queries, obtaining, for one or more of the voice queries, feedback information that references an action taken by a user that submitted the voice query after reviewing a result of the voice query, generating, for the one or more voice queries, a posterior recognition confidence measure that reflects a probability that the voice query was correctly recognized, wherein the posterior recognition confidence measure is generated based at least on the feedback information for the voice query, selecting a subset of the one or more voice queries based on the posterior recognition confidence measures, and adapting an acoustic model using the subset of the voice queries.