Nippon Telegraph and Telephone Corporation

Kiyoaki Matsui

Takaaki Fukutomi

Takashi Nakamura

Ryo Masumura

Manabu Okamoto

Takanobu Oba

Yusuke Shinohara

Yoshikazu Yamaguchi

Takafumi Moriya

An utterance section detection device which is capable of detecting an utterance section with high accuracy on the basis of whether or not an end of a speech section is an end of utterance. The utterance section detection device includes a speech/non-speech determination unit configured to perform speech/non-speech determination which is determination as to whether a certain frame of an acoustic signal is speech or non-speech, an utterance end determination unit configured to perform utterance end determination which is determination as to whether or not an end of a speech section is an end of utterance for each speech section which is a section determined as speech as a result of the speech/non-speech determination, a non-speech section duration threshold determination unit configured to determine a threshold regarding a duration of a non-speech section on the basis of a result of the utterance end determination, and an utterance section detection unit configured to detect an utterance section by comparing a duration of a non-speech section following the speech section with the corresponding threshold.

Utterance section detection device, utterance section detection method, and program

Provided is a technology of learning an acoustic model with a certain degree of accuracy of sound recognition within a short calculation period. An acoustic model learning device includes: a loss calculation unit configured to calculate a loss of sound data which is an element of the corpus Cfor learning by using an acoustic model; a curriculum corpus generation unit configured to generate a curriculum corpus being a union of subsets of the corpuses Cfor learning, the corpuses Cincluding, as elements, sound data for which the loss falls within a predetermined range indicating a small value; an acoustic model update unit configured to update the acoustic model by using the curriculum corpus; and a first end condition determination unit configured to output the acoustic model when a predetermined end condition is satisfied, or transfer execution control to the loss calculation unit when the predetermined end condition is not satisfied, and the acoustic model update unit is configured to update the acoustic model by giving a weight to a gradient for sound data which is an element of the curriculum corpus using such a weight for sound data as to have a smaller value as a number of times the sound data has been selected as an element of the curriculum corpus becomes larger.

Acoustic model learning apparatus, acoustic model learning method, and program

A learning data acquisition device or the like, capable of acquiring learning data by superimposing noise data on clean voice data at an appropriate SN ratio, is provided. The learning data acquisition device includes a voice recognition influence degree calculation unit and a learning data acquisition unit. The voice recognition influence degree calculation unit calculates an influence degree on voice recognition accuracy caused by a change of a signal-to-noise ratio, based on a result of voice recognition on the knoise superimposed voice data and a result of voice recognition on the k−1noise superimposed voice data, where K is an integer of 2 or larger, k=2, 3, . . . , K, and a signal-to-noise ratio of the the knoise superimposed voice data is smaller than a signal-to-noise ratio of the k−1noise superimposed voice data, and obtains a largest signal-to-noise ratio SNRamong signal-to-noise ratios of the k−1noise superimposed voice data when the influence degree meets a given threshold condition. The learning data acquisition unit acquires noise superimposed voice data having a signal-to-noise ratio that is equal to or larger than the signal-to-noise ratio SNR, as learning data.

Learning data acquisition apparatus, model learning apparatus, methods and programs for the same

A voice/non-voice determination device robust with respect to an acoustic signal in a high-noise environment is provided. The voice/non-voice determination device includes an acoustic scene classification unit including a first model which receives input of an acoustic signal and outputs acoustic scene information which is information regarding a scene where the acoustic signal is collected, a speech enhancement unit including a second model which receives input of the acoustic signal and outputs speech enhancement information which is information regarding the acoustic signal after enhancement, and a voice/non-voice determination unit including a third model which receives input of the acoustic signal, the acoustic scene information and the speech enhancement information and outputs a voice/non-voice label which is information regarding a label of either a speech section or a non-speech section.

Voice/non-voice determination device, voice/non-voice determination model parameter learning device, voice/non-voice determination method, voice/non-voice determination model parameter learning method, and program

Detection precision of a non-verbal sound is improved. An acoustic model storage unitA stores an acoustic model that is configured by a deep neural network with a bottleneck structure, and estimates a phoneme state from a sound feature value. A non-verbal sound model storage unitB stores a non-verbal sound model that estimates a posterior probability of a non-verbal sound likeliness from the sound feature value and a bottleneck feature value. A sound feature value extraction unitextracts a sound feature value from an input sound signal. A bottleneck feature value estimation unitinputs the sound feature value to the acoustic model and obtains an output of a bottleneck layer of the acoustic model as a bottleneck feature value. A non-verbal sound detection unitinputs the sound feature value and the bottleneck feature value to the non-verbal sound model and obtains the posterior probability of the non-verbal sound likeliness output by the non-verbal sound model.

Non-verbal utterance detection apparatus, non-verbal utterance detection method, and program

A training speech data generating apparatus includes: a voice conversion unit that converts, using fourth noise data, which is noise data based on third noise data, and speech data, the speech data so as to make the speech data clearly audible under a noise environment corresponding to the fourth noise data; and a noise superimposition unit that obtains training speech data by superimposing the third noise data and the converted speech data.

Growing community of inventors

Tokyo, Japan

Kiyoaki Matsui