Nec Corporation

University of Washington

Makoto Terao

Takafumi Koshinaka

Shuhei Yoshida

Yoshifumi Onishi

Masahiro Tani

Koji Okabe

Katsuhiko Takahashi

Soma Shiraishi

Takami Sato

Shinichi Ando

Yu Nabeto

Ryosuke Sakai

Tomokazu Kaneko

Mari Ostendorf

A recognition loss calculation unit of a learning device calculates a recognition loss using: a recognition result with respect to recognition object data in a learning data set that is a set of a pair of the recognition object data and a weak label; a mixing matrix calculated based on the learning data set; and the weak label attached to the recognition object data. The recognition loss calculation unit includes: a difference calculation unit that calculates a difference between a mixing matrix and the recognition result; and a sum of squares calculation unit that calculates the recognition loss by calculating a sum of a square of the difference.

Learning device, learning method, and storage medium for learning device

An image acquisition unitacquires a plurality of images. The plurality of images include an object to be inferred. An image cut-out unitcuts out an object region including the object from each of the plurality of images acquired by the image acquisition unit. An importance generation unitgenerates importance information by processing the object region cut out by the image cut-out unit. The importance information indicates the importance of the object region when an object inference model is generated, and is generated for each object region, that is, for each image acquired by the image acquisition unit. A learning data generation unitstores a plurality of object regions cut out by the image cut-out unitand a plurality of pieces of importance information generated by the importance generation unitin a learning data storage unitas at least a part of the learning data.

Generating learning data from important, cut-out object regions

The dataset supply unit supplies a learning dataset. The recognition unit outputs the recognition result for the recognition object data in the supplied learning dataset. Further, the intersection matrix computation unit computes the intersection matrix based on the learning dataset. The recognition loss computation unit computes the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data. Then, the updating unit updates the parameters of the recognition unit based on the recognition loss.

Learning device, learning method and recording medium

The learning deviceD is learned to extract moving image feature amount Fm which is feature amount relating to the moving image data Dm when the moving image data Dm is inputted thereto, and is learned to extract still image feature amount Fs which is feature amount relating to the still image data Ds when the still image data Ds is inputted thereto. The first inference unitD performs a first inference regarding the moving image data Dm based on the moving image feature amount Fm. The second inference unitD performs a second inference regarding the still image data Ds based on the still image feature amount Fs. The learning unitD performs learning of the feature extraction unitD based on the results of the first inference and the second inference.

Updated learning of feature extraction model that extracts moving image feature amount from moving image data and still image feature amount from still image data

A model learning device provided with: an error-added movement locus generation unit for adding an error to movement locus data for action learning that represents the movement locus of a subject and to which is assigned an action label that is information representing the action of the subject, and thereby generating error-added movement locus data; and an action recognition model learning unit for learning a model, using at least the error-added movement locus data and learning data created on the basis of the action label, by which model the action of some subject can be recognized from the movement locus of the subject. Thus, it is possible to provide a model by which the action of a subject can be recognized with high accuracy on the basis of the movement locus of the subject estimated using a camera image.

Model learning device, model learning method, and recording medium

An analysis object determination device includes a detection unit which detects a plurality of specific utterance sections using data related to a voice in a conversation, the specific utterance sections representing a plurality of specific events originating from one or a plurality of participants in the conversation, or a specific event originating from one of the conversation participants, and an object determination unit which determines, on the basis of the plurality of specific utterance sections detected by the detection unit, one or more cause analysis sections for the specific event originating from the conversation participant, the number of the cause analysis sections being fewer than the number of the plurality of specific utterance sections.

Analysis object determination device, analysis object determination method and computer-readable medium

An analysis subject determination device includes: a demand period detection unit which detects, from data corresponding to audio of a dissatisfaction conversation, a demand utterance period which represents a demand utterance of a first conversation party among a plurality of conversation parties which are carrying out the dissatisfaction conversation; a negation period detection unit which detects, from the data, a negation utterance period which represents a negation utterance of a second conversation party which differs from the first conversation party; and a subject determination unit which, from the data, determines a period with a time obtained from the demand period utterance period as a start point and a time obtained from the negation utterance period after the demand utterance period as an end point to be an analysis subject period of a cause of dissatisfaction of the first conversation party in the dissatisfaction conversation.

Analysis object determination device and analysis object determination method

An apparatus that calculates a confidence measure of a target word string specified in a recognition result includes: an alternative candidate generator which generates an alternative candidate word string in the position of the target word string; a classifier training unit which trains a classifier which is configured to discriminate between the target word string and the alternative candidate word string; a feature extractor which extracts a feature value representing an adjacent context in the position of the target word string; and a confidence measure calculator which determining whether the true word string in the position of the target word string is the target word string or the alternative candidate word string by using the classifier and the feature value and calculates a confidence measure of the target word string on the basis of the determination result.

Relative semantic confidence measure for error detection in ASR

A sound segment sorting unit () sorts the sound segments of a video. An image segment sorting unit () sorts the image segments of the video. A multiple sorting result generation unit () generates a plurality of sound segment sorting results and/or a plurality of image segment sorting results. A sorting result pair generation unit () generates a plurality of sorting result pairs of the sorting results as the candidates of the optimum segment sorting result of the video. A sorting result output unit () compares the sorting result comparative scores of the sorting result pairs calculated by a sorting result comparative score calculation unit () and thus outputs a sound segment sorting result and an image segment sorting result having good correspondence. This allows to accurately sort, for each object, a plurality of sound segments and a plurality of image segments contained in the video without adjusting parameters in advance.

Sound and image segment sorting device and method

There is provided an apparatus including a model based topic segmentation section that segments a text using a topic model representing semantic coherence, a parameter estimation section that estimates a control parameter used in segmenting the text based on detection of a change point of word distribution in the text, using the result of segmentation by the model based topic segmentation unit as training data, and a change point detection topic segmentation section that segments the text, based on detection of the change point of word distribution in the text, using the parameter estimated by the parameter estimation section.

Apparatus, method and program for text segmentation

A voice judging system including feature value extraction means that analyzes a sound signal input from a sound signal input device, and extracts a time series of the feature values, sub-word boundary score calculating means that calculates a time series of sub-word boundary scores, by having reference to sound models of voice stored in a voice model storage unit, temporal regularity analyzing means that analyzes temporal regularity of the sub-word boundary scores, and voice judgment means judges whether the input sound signal is voice or non-voice using of the temporal regularity of the sub-word boundary scores.

Voice judging system, voice judging method and program for voice judgment

[PROBLEMS] To provide a speech-to-text system and the like capable of matching edit result text acquired by editing recognition result text or edit result text which is newly-written text information with speech data. [MEANS FOR SOLVING PROBLEMS] A speech-to-text system () includes a matching unit () which collates edit result text acquired by a text editor unit () with speech recognition result information having time information created by a speech recognition unit () to thereby match the edit result text and speech data.

Speech-to-text system, speech-to-text method, and speech-to-text program

[PROBLEMS] To accurately calculate similarity between media data and a query even if the media data or its meta data has an error. [MEANS FOR SOLVING THE PROBLEMS] A similarity calculation device includes: a single score calculation device used when calculating similarity between first media data and a query, which calculates a single score that shows similarity between second media data different from the first media data and the query; an inter-media similarity calculation device which calculates inter-media similarity that shows the similarity between the second media data and the first media data; and a query similarity calculation device which obtains similarity between the first media data and the query by using the inter-media similarity of the second media data and the single score.

Growing community of inventors

Tokyo, Japan

Makoto Terao