Electronics and Telecommunications Research Institute

Byung Ok Kang

Jeon Gue Park

Hoon Chung

Sung Joo Lee

Yun Keun Lee

Jeom Ja Kang

Ho-Young Jung

Hyung-Bae Jeon

Ho Young Jung

Ji Hyun Wang

Jong Jin Kim

Eui Sok Chung

Yunkeun Lee

Hwa Jeon Song

Ki-young Park

Euisok Chung

Hyung Bae Jeon

Kap Kee Kim

Ki Young Park

Yoonhyung Kim

Hyun Woo Kim

Yoo Rhee Oh

Hoon Young Cho

Hoon-Young Cho

Provided is a self-supervised learning method based on permutation invariant cross entropy. A self-supervised learning method based on permutation invariant cross entropy performed by an electronic device includes: defining a cross entropy loss function for pre-training of an end-to-end speech recognition model; configuring non-transcription speech corpus data composed only of speech as input data of the cross entropy loss function; setting all permutations of classes included in the non-transcription speech corpus data as an output target and calculating cross entropy losses for each class; and determining a minimum cross entropy loss among the calculated cross entropy losses for each class as a final loss.

Self-supervised learning method based on permutation invariant cross entropy and electronic device thereof

Disclosed herein are an apparatus and method for audio-video sampling frequency ratio unification, including memory configured to store at least one program, and a processor configured to execute the program, wherein the program is configured to perform receiving an audio signal and a video signal, adjusting a ratio of a sampling frequency of the audio signal to a sampling frequency of the video signal so that the sampling frequency ratio is constant based on a deep learning network, and outputting an adjusted audio signal and the video signal.

Apparatus and method for audio-video sampling frequency ratio unification

Provided are sentence embedding method and apparatus based on subword embedding and skip-thoughts. To integrate skip-thought sentence embedding learning methodology with a subword embedding technique, a skip-thought sentence embedding learning method based on subword embedding and methodology for simultaneously learning subword embedding learning and skip-thought sentence embedding learning, that is, multitask learning methodology, are provided as methodology for applying intra-sentence contextual information to subword embedding in the case of subword embedding learning. This makes it possible to apply a sentence embedding approach to agglutinative languages such as Korean in a bag-of-words form. Also, skip-thought sentence embedding learning methodology is integrated with a subword embedding technique such that intra-sentence contextual information can be used in the case of subword embedding learning. A proposed model minimizes additional training parameters based on sentence embedding such that most training results may be accumulated in a subword embedding parameter.

Sentence embedding method and apparatus based on subword embedding and skip-thoughts

Provided is a method of automatically expanding input text. The method includes receiving input text composed of a plurality of documents, extracting a sentence pair that is present in different documents among the plurality of documents, setting the extracted sentence pair as an input of an encoder of a sequence-to-sequence model, setting an output of the encoder as an output of a decoder of the sequence-to-sequence model and generating a sentence corresponding to the input, and generating expanded text based on the generated sentence.

System and method for automatically expanding input text

A speech recognition apparatus based on a deep-neural-network (DNN) sound model includes a memory and a processor. As the processor executes a program stored in the memory, the processor generates sound-model state sets corresponding to a plurality of pieces of set training speech data included in multi-set training speech data, generates a multi-set state cluster from the sound-model state sets, and sets the multi-set training speech data as an input node and the multi-set state cluster as output nodes so as to learn a DNN structured parameter.

Apparatus and method for recognizing speech based on a deep-neural-network (DNN) sound model

Provided is an apparatus for large vocabulary continuous speech recognition (LVCSR) based on a context-dependent deep neural network hidden Markov model (CD-DNN-HMM) algorithm. The apparatus may include an extractor configured to extract acoustic model-state level information corresponding to an input speech signal from a training data model set using at least one of a first feature vector based on a gammatone filterbank signal analysis algorithm and a second feature vector based on a bottleneck algorithm, and a speech recognizer configured to provide a result of recognizing the input speech signal based on the extracted acoustic model-state level information.

Apparatus and method for large vocabulary continuous speech recognition

Method of the present invention may include receiving speech feature vector converted from speech signal, performing first search by applying first language model to the received speech feature vector, and outputting word lattice and first acoustic score of the word lattice as continuous speech recognition result, outputting second acoustic score as phoneme recognition result by applying an acoustic model to the speech feature vector, comparing the first acoustic score of the continuous speech recognition result with the second acoustic score of the phoneme recognition result, outputting first language model weight when the first coustic score of the continuous speech recognition result is better than the second acoustic score of the phoneme recognition result and performing a second search by applying a second language model weight, which is the same as the output first language model, to the word lattice.

Method for estimating language model weight and system for the same

A speech recognition system includes: a speed level classifier for measuring a moving speed of a moving object by using a noise signal at an initial time of speech recognition to determine a speed level of the moving object; a first speech enhancement unit for enhancing sound quality of an input speech signal of the speech recognition by using a Wiener filter, if the speed level of the moving object is equal to or lower than a specific level; and a second speech enhancement unit enhancing the sound quality of the input speech signal by using a Gaussian mixture model, if the speed level of the moving object is higher than the specific level. The system further includes an end point detection unit for detecting start and end points, an elimination unit for eliminating sudden noise components based on a sudden noise Gaussian mixture model.

Noise reduction for speech recognition in a moving vehicle

An utterance verification method for an isolated word N-best speech recognition result includes: calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model based on an N-best speech recognition result for an input utterance; measuring a confidence score of an N-best speech-recognized word using the log likelihoods; calculating distance between phonemes for the N-best speech-recognized word; comparing the confidence score with a threshold and the distance with a predetermined mean of distances; and accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptance.

Utterance verification method and apparatus for isolated word N-best recognition result

A method for separating a sound source from a mixed signal, includes Transforming a mixed signal to channel signals in frequency domain; and grouping several frequency bands for each channel signal to form frequency clusters. Further, the method for separating the sound source from the mixed signal includes separating the frequency clusters by applying a blind source separation to signals in frequency domain for each frequency cluster; and integrating the spectrums of the separated signal to restore the sound source in a time domain wherein each of the separated signals expresses one sound source.

Method for separating source signals and apparatus thereof

A model-based distortion compensating noise reduction apparatus for speech recognition, includes: a speech absence probability calculator for calculating the probability distribution for absence and existence of a speech using the sound absence and existence information for the frames; a noise estimation updater for estimating a more accurate noise component by updating the variance of the clean speech and noise for each frame; and a speech absence probability-based noise filter for outputting a first clean speech through the speech absence probability transmitted from the speech absence probability calculator and a first noise filter. Further, the model-based distortion compensating noise reduction apparatus includes a post probability calculator for calculating post probabilities for mixtures using a GMM containing a clean speech in the first clean speech; and a final filter designer for forming a second noise filter and outputting an improved final clean speech signal using the second noise filter.

Model-based distortion compensating noise reduction apparatus and method for speech recognition

A Viterbi decoder includes: an observation vector sequence generator for generating an observation vector sequence by converting an input speech to a sequence of observation vectors; a local optimal state calculator for obtaining a partial state sequence having a maximum similarity up to a current observation vector as an optimal state; an observation probability calculator for obtaining, as a current observation probability, a probability for observing the current observation vector in the optimal state; a buffer for storing therein a specific number of previous observation probabilities; a non-linear filter for calculating a filtered probability by using the previous observation probabilities stored in the buffer and the current observation probability; and a maximum likelihood calculator for calculating a partial maximum likelihood by using the filtered probability. The filtered probability may be a maximum value, a mean value or a median value of the previous observation probabilities and the current observation probability.

Viterbi decoder and speech recognition method using same using non-linear filter for observation probabilities

A noise cancellation apparatus includes a noise estimation module for receiving a noise-containing input speech, and estimating a noise therefrom to output the estimated noise; a first Wiener filter module for receiving the input speech, and applying a first Wiener filter thereto to output a first estimation of clean speech; a database for storing data of a Gaussian mixture model for modeling clean speech; and an MMSE estimation module for receiving the first estimation of clean speech and the data of the Gaussian mixture model to output a second estimation of clean speech. The apparatus further includes a final clean speech estimation module for receiving the second estimation of clean speech from the MMSE estimation module and the estimated noise from the noise estimation module, and obtaining a final Wiener filter gain therefrom to output a final estimation of clean speech by applying the final Wiener filter gain.

Noise cancellation system and method

A microphone-array-based speech recognition system using a blind source separation (BBS) and a target speech extraction method in the system are provided. The speech recognition system performs an independent component analysis (ICA) to separate mixed signals input through a plurality of microphone into sound-source signals, extracts one target speech spoken for speech recognition from the separated sound-source signals by using a Gaussian mixture model (GMM) or a hidden Markov Model (HMM), and automatically recognizes a desired speech from the extracted target speech. Accordingly, it is possible to obtain a high speech recognition rate even in a noise environment.

Microphone array based speech recognition system and target speech extracting method of the system

Provided are an apparatus and method for generating a noise adaptive acoustic model including a noise adaptive discriminative adaptation method. The method includes: generating a baseline model parameter from large-capacity speech training data including various noise environments; and receiving the generated baseline model parameter and applying a discriminative adaptation method to the generated results to generate an migrated acoustic model parameter suitable for an actually applied environment.

Growing community of inventors

Daejeon, South Korea

Byung Ok Kang