• Àüü
  • ÀüÀÚ/Àü±â
  • Åë½Å
  • ÄÄÇ»ÅÍ
´Ý±â

»çÀÌÆ®¸Ê

Loading..

Please wait....

±¹³» ÇÐȸÁö

Ȩ Ȩ > ¿¬±¸¹®Çå > ±¹³» ÇÐȸÁö > µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

Current Result Document :

ÇѱÛÁ¦¸ñ(Korean Title) Ãæ½Çµµ ³ôÀº ±âħ ¼Ò¸® »ý¼ºÀ» À§ÇÑ °³¼±µÈ GAN
¿µ¹®Á¦¸ñ(English Title) An improved GAN for generating high-fidelity synthesized cough sounds
ÀúÀÚ(Author) ¹é¹®±â   À̱Ôö   Moon-Ki Back   Kyu-Chul Lee  
¿ø¹®¼ö·Ïó(Citation) VOL 36 NO. 03 PP. 0036 ~ 0055 (2020. 12)
Çѱ۳»¿ë
(Korean Abstract)
GAN(Generative Adversarial Network)Àº ÄÄÇ»ÅÍ ºñÀü ºÐ¾ß¿¡¼­ Å« Àα⸦ ¾ò¾ú°í, À̹ÌÁö »ý¼º ÀÛ¾÷¿¡ ³Î¸® »ç¿ëµÇ°í ÀÖ´Ù. ±×¸®°í ÃÖ±Ù, GAN ¿¬±¸ÀÚµéÀº GANÀ» »ç¿ëÇÏ¿© ¼Ò¸® µ¥ÀÌÅ͸¦ »ý¼ºÇϱ⠽ÃÀÛÇß´Ù. ÆÄÇü Àº À̹ÌÁö¿Í ´Ù¸£°Ô ÀÌ»ê °ªÀ¸·Î ±¸¼ºµÈ ½ÅÈ£À̹ǷÎ, À̹ÌÁö ÇнÀ¿¡ ÁÖ·Î »ç¿ëÇÏ´Â CNN(Convolutional Neural Network)À» È°¿ëÇÏ¿© ÆÄÇüÀ» ÇнÀÇϱ⠾î·Æ´Ù. À̸¦ ±Øº¹Çϱâ À§ÇÏ¿©, GAN ¿¬±¸ÀÚµéÀº ±âÁ¸ À̹ÌÁö »ý¼º GANÀ» Àç»ç¿ëÇÏ¿© ÆÄÇü ´ë½Å ½Ã°£-ÁÖÆļö Ç¥ÇöÀ» ÇнÀÇÏ´Â Á¢±ÙÀ» Á¦¾ÈÇß´Ù. ÀÌ·¯ÇÑ Á¢±ÙÀ» µû¶ó¼­, º» ³í¹®Àº »ý¼ºµÈ ÆÄÇüÀÇ Ãæ½Çµµ(fidelity)¸¦ °³¼±Çϱâ À§ÇÑ °³¼±µÈ ¼Ò¸® »ý¼º GANÀ» Á¦¾ÈÇÑ´Ù. °³¼±µÈ GAN Àº HPSS(Harmonic Percussive Source Separation)¸¦ »ç¿ëÇØ ½Ã°£¿¡ µû¸¥ ½ºÆåÆ®·³ÀÇ Æ¯Â¡À» ÃßÃâÇÏ°í, Á¡Áø ÀûÀ¸·Î ¼ºÀåÇÏ´Â ³×Æ®¿öÅ©¸¦ ÅëÇØ »ý¼ºµÇ´Â ÆÄÇüÀÇ Ç°ÁúÀ» °³¼±Çϴ Ư¡ÀÌ ÀÖ´Ù. º» ³í¹®¿¡¼­´Â °ø°³µÈ ±âħ µ¥ÀÌÅͼ¼Æ®¸¦ »ç¿ëÇØ Á¦¾ÈÇÑ GANÀ» ÇнÀ½ÃÅ°°í, Ãæ½Çµµ¿Í ´Ù¾ç¼º(diversity) Ãø¸é¿¡¼­ ¼º´ÉÀ» Æò°¡ÇÑ´Ù.
¿µ¹®³»¿ë
(English Abstract)
Generative Adversarial Networks (GANs) have gained tremendous popularity in computer vision, and have been widely used for image generation tasks. Recently, GAN researchers have started gen- erating sound data by using GANs. Unlike images, a waveform is a sampled signal consisting of discrete samples, so it is not easy to learn the waveform by utilizing Convolutional Neural Network (CNN), which is mainly trained on natural images. To overcome this difficulty, GAN researchers proposed an approach employing time-frequency representations instead of time-series waveforms to reuse existing image-generating GANs. Following this approach, we propose an improved sound-generating GAN to improve the fidelity of generated waveforms. We designed a network that first uses Harmonic Percussive Source Separation (HPSS) to extract spectral features over time and then improves the quality of generated waveforms by applying progressively-growing networks. In this paper, we train our GAN on a public cough dataset and evaluate the perform- ances in terms of the fidelity and diversity of generated waveforms
Å°¿öµå(Keyword) »ý¼ºÀû Àû´ë ½Å°æ¸Á   »ý¼º ¸ðµ¨   ¼Ò¸® µ¥ÀÌÅÍ   ±âħ ¼Ò¸®   ³ôÀº Ãæ½Çµµ   Generative Adversarial Network   generative model   sound data   cough sound   high-fidelity  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå