Abstract: Provided is a voice search technology that can efficiently find and check a problematic call. To this end, a voice search system of the present invention includes a call search database that stores, for each of a reception channel and a transmission channel of each of a plurality of pieces of recorded call voice data, voice section sequences in association with predetermined keywords and time information. The call search database is searched based on an input search keyword, so that a voice section sequence that contains the search keyword is obtained.

Abstract: [Object] An object is to provide a training method of improving training of a recurrent neural network (RNN) using time-sequential data. [Solution] The training method includes a step 220 of initializing the RNN, and a training step 226 of training the RNN by designating a certain vector as a start position and optimizing various parameters to minimize error function. The training step 226 includes: an updating step 250 of updating RNN parameters through Truncated BPTT using consecutive N (N?3) vectors having a designated vector as a start point and using a reference value of a tail vector as a correct label; and a first repetition step 240 of repeating the process of executing the training step by newly designating a vector at a position satisfying a prescribed relation with the tail of N vectors used at the updating step until an end condition is satisfied. The vector at a position satisfying the prescribed relation is positioned at least two vectors behind the designated vector.

Abstract: A speaker estimation method that estimate the speaker from audio and image includes: inputting audio; extracting a feature quantity representing a voice characteristic from the input audio; inputting an image; detecting person regions of respective persons from the input image; estimating feature quantities representing voice characteristics from the respective detected person regions; Performing a change such that an image taken from another position and with another angle is input when any person is not detected; calculating a similarity between the feature quantity representing the voice characteristic extracted from the audio and the feature quantity representing the voice characteristic estimated from the person region in the image; and estimating a speaker from the calculated similarity.

Abstract: A speech recognition device includes: an acoustic model based on an End-to-End neural network responsive to an observed sequence formed of prescribed acoustic features obtained from a speech signal by feature extracting unit, for calculating probability of the observed sequence being a certain symbol sequence; and a decoder responsive to a symbol sequence candidate, for decoding a speech signal by a WFST based on a posterior probability of each of word sequences corresponding to the symbol sequence candidate, probabilities calculated by the acoustic model for symbol sequences selected based on an observed sequence, and a posterior probability of each of the plurality of symbol sequences.

Abstract: [Object] An object is to provide a speech recognition device with improved recognition accuracy using characteristics of a neural network. [Solution] A speech recognition device includes: an acoustic model 308 implemented by a RNN (recurrent neural network) for calculating, for each state sequence, the posterior probability of a state sequence in response to an observed sequence consisting of prescribed speech features obtained from a speech; a WFST 320 based on S?1HCLG calculating, for each word sequence, posterior probability of a word sequence in response to a state sequence; and a hypothesis selecting unit 322, performing speech recognition of the speech signal based on a score calculated for each hypothesis of a word sequence corresponding to the speech signal, using the posterior probabilities calculated by the acoustic model 308 and the WFST 320 for the input observed sequence.

Abstract: The present invention pertains to a method for estimating a sound source position in a space with high accuracy using a microphone installed on a robot moving in the space. A mobile robot includes a self-position estimation unit configure to estimate the self-position of the mobile robot, a sound source information obtaining unit configured to obtain direction information of an observed sound source, and a sound source position estimation unit configured to estimate the position of the sound source based on the estimated self-position and the direction information of the sound source.

Abstract: [Object] An object is to provide a training method of improving training of a recurrent neural network (RNN) using time-sequential data. [Solution] The training method includes a step 220 of initializing the RNN, and a training step 226 of training the RNN by designating a certain vector as a start position and optimizing various parameters to minimize error function. The training step 226 includes: an updating step 250 of updating RNN parameters through Truncated BPTT using consecutive N (N?3) vectors having a designated vector as a start point and using a reference value of a tail vector as a correct label; and a first repetition step 240 of repeating the process of executing the training step by newly designating a vector at a position satisfying a prescribed relation with the tail of N vectors used at the updating step until an end condition is satisfied. The vector at a position satisfying the prescribed relation is positioned at least two vectors behind the designated vector.

Abstract: Provided is a voice search technology that can efficiently find and check a problematic call. To this end, a voice search system of the present invention includes a call search database that stores, for each of a reception channel and a transmission channel of each of a plurality of pieces of recorded call voice data, voice section sequences in association with predetermined keywords and time information. The call search database is searched based on an input search keyword, so that a voice section sequence that contains the search keyword is obtained.

Abstract: The present invention pertains to a method for estimating a sound source position in a space with high accuracy using a microphone installed on a robot moving in the space. A mobile robot includes a self-position estimation unit configure to estimate the self-position of the mobile robot, a sound source information obtaining unit configured to obtain direction information of an observed sound source, and a sound source position estimation unit configured to estimate the position of the sound source based on the estimated self-position and the direction information of the sound source.

Abstract: An acoustic feature representing speech data provided with meta data is extracted. Next, a group of acoustic features which are extracted only from the speech data containing a specific word in the meta data and not from the other speech data is extracted from obtained sub-groups of acoustic features. The word and the extracted group of acoustic features are associated with each other to be stored. When there is a search key matching the word in the input search keys, the group of acoustic features corresponding to the word is output. Accordingly, the efforts of a user for inputting a key when the user searches for speech data are reduced.