Resume

Emotion recognition (2017 ~ )

- Non-speech Emotion Recognition (e.g. laugh, sigh, scream, yawn, cry, etc)

- Speech Emotion Recognition (e.g. anger, boredom, disgust, fear, happiness, neutral, sadness)

- Find and classify functions with continuous functions, including important pitch related functions, formants, energy-related functions and timing functions

- Transfer learning to emotion audio set with massive youtube audio set using VGG model

- Using CRNN(CNN-RNN), BiLSTM, Self Attention model

- Overfitting prevent Method( speaker independent-cross validation, batch normalization layer, early stopping, l2 using regularization loss, dropout)

- The OpenSMILE toolkit is utilized to extract audio feature sets, including eGemaps, IS09, IS10, IS11, IS13, MFCC, logMel features

- Emotion recognition of speech and image fusion using Attention model - Analysis of emotion recognition data based on LSTM using Influence function

 

• Speech recognition (2017 ~ )

- Implement real-time speech recognition with English and Korean version using Kaldi toolkit - Using Ubuntu environment shell script and python language

- The English speech recognition version used WSJ(Wall Street Journal corpus) and LDC93S6B dataset

- The Korean speech recognition version used approximately 600 hours of Database ( ETRI 2002, ETRI 2003, ETRI 2011 SMART MOBILE, SITEC CAR, PBW452, data augmentation, self dataset, etc)

- In the acoustic model, speech features of MFCC(Mel Frequency Cepstral Coefficient) are extracted

- In the language model, tri-gram model is used - Generate HCLG. FST file by fusion of language model and acoustic model using WFST based

- Learn using deep learning network

​

• Source localization (2017 ~ )

- Implementation of sound source localization using GCC-PATH - Using four horizontal microphone arrays

- Audio signals cross-correlation when entering the four microphones

​

• Speech Separation with GCC-NMF (2017)

- Source separation algorithm using GCC-NMF that combines unsupervised dictionary learning vas nonnegative matrix factorization (NMF) with spatial localization via the generalized cross correlation method

- Using publicly available data from the SISEC signal separation evaluation campaign

- Using python language

​

• Rule-Based Chatbot (2017 )

- Using RiveScript artificial intelligence script language in python

- Organize a scenario a hamburger shop, weather content, daily conversation in Korean version

- Configure server and client in python and C language with network socket