Nec Corporation

Nec Informatec Systems, Ltd.

Tadashi Emori

Yoshifumi Onishi

Masanori Tsujikawa

Ryosuke Isotani

Kazuhiro Arai

Masahiro Tani

Koichi Shinoda

A text in a corpus including a set of world wide web (web) pages is analyzed. At least one word appropriate for a document type set according to a voice recognition target is extracted based on an analysis result. A word set is generated from the extracted at least one word. A retrieval engine is caused to perform a retrieval process using the generated word set as a retrieval query of the retrieval engine on the Internet, and a link to a web page from the retrieval result is acquired. A language model for voice recognition is generated from the acquired web page.

Language model generating device, method thereof, and recording medium storing program thereof

A method for processing multichannel acoustic signals which is characterized by calculating the feature quantity of each channel from the input signals of a plurality of channels, calculating similarity between the channels in the feature quantity of each channel, selecting channels having high similarity, and separating signals using the input signals of the selected channels.

Method for processing multichannel acoustic signal, system therefor, and program

A method for processing multichannel acoustic signals which processes input signals of a plurality of channels including the voices of a plurality of speaking persons. The method is characterized by detecting the voice section of each speaking person or each channel, detecting overlapped sections wherein the detected voice sections are common between channels, determining a channel to be subjected to crosstalk removal and the section thereof by use of at least voice sections not including the detected overlapped sections, and removing crosstalk in the sections of the channel to be subjected to the crosstalk removal.

A method for processing multichannel acoustic signals, whereby input signals of a plurality of channels including the voices of a plurality of speaking persons are processed. The method is characterized by comprising: calculating the first feature quantity of the input signals of the multichannels for each channel; calculating similarity of the first feature quantity of each channel between the channels; selecting channels having high similarity; separating signals using the input signals of the selected channels; inputting the input signals of the channels having low similarity and the signals after the signal separation; and detecting a voice section of each speaking person or each channel.

Method for processing multichannel acoustic signal, system thereof, and program

A warping factor estimation system comprises label information generation unit that outputs voice/non-voice label information, warp model storage unit in which a probability model representing voice and non-voice occurrence probabilities is stored, and warp estimation unit that calculates a warping factor in the frequency axis direction using the probability model representing voice and non-voice occurrence probabilities, voice and non-voice labels, and a cepstrum.

Frequency axis warping factor estimation apparatus, system, method and program

A speech sound detection apparatus receives an input audio signal (as a sound reception unit), and computes input power that indicates a magnitude of the sound represented by the audio signal (as an input power computation unit). The apparatus estimates a correction function that is a continuous function defining a relation between a certain frequency and a correction coefficient used to approximate the input power computed at that frequency to the reference power predetermined for that frequency (as a correction function estimation unit). The apparatus corrects the input power at every frequency, based upon the correction coefficient that is obtained in accordance with the relation defined by the estimated correction function (as an input power correcting unit). The apparatus further determines whether or not the sound represented by the received audio signal is speech sound, based upon the corrected input power (as a speech sound detection unit).

Speech sound detection apparatus

A signal correction apparatus receives an input audio signal (serving as a first sound reception means). The signal correction apparatus computes, at every frequency, first power that indicates magnitude of sound represented by the input audio signal (serving as a first power computation means). The signal correction apparatus estimates a correction function that is a continuous function defining a relation between each frequency and a correction coefficient used to approximate the first power computed at that frequency to the reference power predetermined for that frequency (serving as a correction function estimation means). The signal correction apparatus multiplies the computed first power by the correction coefficient acquired in accordance with the relation defined by the estimated correction function so as to correct the first power at every frequency (serving as a power correcting means).

Signal correction apparatus equipped with correction function estimation unit

A language model learning system for learning a language model on an identifiable basis relating to a word error rate used in speech recognition. The language model learning system () includes a recognizing device () for recognizing an input speech by using a sound model and a language model and outputting the recognized word sequence as the recognition result, a reliability degree computing device () for computing the degree of reliability of the word sequence, and a language model parameter updating device () for updating the parameters of the language model by using the degree of reliability. The language model parameter updating device updates the parameters of the language model to heighten the degree of reliability of the word sequence the computed degree of reliability of which is low when the recognizing device recognizes by using the updated language model and the reliability degree computing device computes the degree of reliability.

Language model learning system, language model learning method, and language model learning program

A speech signal processing device is equipped with a power acquisition unit, a probability distribution acquisition unit, and a correspondence degree determination unit. The power acquisition unit accepts an inputted speech signal and, based on the accepted speech signal, acquires power representing the intensity of a speech sound represented by the speech signal. The probability distribution acquisition unit acquires a probability distribution using the intensity of the power acquired by the power acquisition unit as a random variable. The correspondence degree determination unit determines whether a correspondence degree representing a degree that power acquired by the power acquisition unit in a case that a predetermined reference speech signal is inputted into the power acquisition unit corresponds with predetermined reference power is higher than a predetermined reference correspondence degree, based on the probability distribution acquired by the probability distribution acquisition unit.

Speech signal processing device

Acoustic models and language models are learned according to a speaking length which indicates a length of a speaking section in speech data, and speech recognition process is implemented by using the learned acoustic models and language models. A speech recognition apparatus includes means () for detecting a speaking section in speech data () and for generating a section information which indicates the detected speaking section, means () for recognizing a data part corresponding to a section information in the speech data as well as text data () written from the speech data and for classifying the data part based on a speaking length thereof, and means () for learning acoustic models and language models () by using the classified data part ().

Audio recognition apparatus and speech recognition method using acoustic models and language models

To this end, a voice detection device includes a band-based power calculation unit that calculates a total of signal power values (sub-band power) of signals entered from the microphones from one preset frequency width (sub-band) to another. The voice detection device also includes a band-based noise estimation unit that estimates the sub-band based noise power, and a sub-band based SNR calculation unit. The sub-band based SNR calculation unit calculates a sub-band SNR from one sub-band to another to output the largest one of the sub-band SNRs as an SNR for a microphone of interest. The voice detection device further includes a voice/non-voice decision unit that determines the voice/non-voice using the SNR for the microphone of interest.

Device, method and program for voice detection and recording medium

A weighting factor learning system includes an audio recognition section that recognizes learning audio data and outputting the recognition result; a weighting factor updating section that updates a weighting factor applied to a score obtained from an acoustic model and a language model so that the difference between a correct-answer score calculated with the use of a correct-answer text of the learning audio data and a score of the recognition result becomes large; a convergence determination section that determines, with the use of the score after updating, whether to return to the weighting factor updating section to update the weighting factor again; and a weighting factor convergence determination section that determines, with the use of the score after updating, whether to return to the audio recognition section to perform the process again and update the weighting factor using the weighting factor updating section.

Weighting factor learning system and audio recognition system

To enable selection of a speaker, the acoustic feature value of which is similar to that of an utterance speaker, with accuracy and stability, while adapting to changes even when the acoustic feature value of the speaker changes every moment, a long-time speaker score is calculated (log likelihood of each of a plurality of speaker models stored in a speaker model storage with respect to the acoustic feature value) based on an arbitrary number of utterances, for example, and a short-time speaker score is calculated based on a short-time utterance, for example. Speakers are selected corresponding to a predetermined number of speaker models having a high long-time speaker score. Speakers are selected corresponding to the speaker models, the number of which is smaller than the predetermined number and the short-time speaker sore of which is high, from among the speakers having a high long-time speaker score.

Speaker selection based at least on an acoustic feature value similar to that of an utterance speaker

A speech recognition apparatus () selects an optimum recognition result from recognition results output from a set of speech recognizers (s-sM) based on a majority decision. This decision is implemented with taking into account weight values, as to the set of the speech recognizers, learned by a learning apparatus (). The learning apparatus includes a unit () selecting speech recognizers corresponding to characteristics of speech for learning (), a unit () finding recognition results of the speech for learning by using the selected speech recognizers, a unit () unifying the recognition results and generating a word string network, and a unit () finding weight values concerning a set of the speech recognizers by implementing learning processing. When finding weight values, the learning apparatus selects a word from each arc set in the word string network based on a majority decision which is taken into account candidates of weight value, and outputs weight value candidates which minimize a recognition error rate of a word string formed of the selected words, as a learning result.

Growing community of inventors

Tokyo, Japan

Tadashi Emori