20150331 msr outreach media_roundtable_deck_연세대강홍구교수_음성합성

  • Published on
    26-Jul-2015

  • View
    505

  • Download
    5

Embed Size (px)

Transcript

<p> 1. 2. 21 3. 4. / () () 1472 - / - (HCI) 5. 7 Speech (: Dr. Frank Soong) 2008~09 : / 2010~11 : 2011~12 : - 2012~13 : HMM TTS 2013~14 : TTS 2014~ : TTS : / Speech group 6. - 1 - 2 1 ( , ) - 3 on-site 1:1 () (6) - 7 () (3) - 5 2 ( 6, 1, 3) 7. Speech Group Korean Day 2013 Korean day Korean day 8. , , 9. / 10. ( 3) - (Text-to-Speech) 11. , , , : , , : , : , , / - (TTS) (1) 12. , - (TTS) (2) - TTS &amp; &amp; 13. - (TTS) (3) ) Microsoft Cortana ( , ) Pepper [] Get Started with Cortana - http://youtube.com/watch?v=tQFrd6SEiLM 14. : [DECtalk] , [AT&amp;T] [HTS] - (TTS) (4) [] DECtalk : http://www.speechfxinc.com/dectalk.html AT&amp;T : http://www2.research.att.com/~ttsweb/tts/demo.php HTS : http://hts.sp.nitech.ac.jp/nitech-naist-hts_blizzard2006 15. (source)-(filter) - : : (excitation) - (TTS) (5) )n(s)z(A1 1 e(n) gain impulse train random noise pitch V UV . 16. (Human-Computer), (Human-Human) (Speech synthesis, Text-to-Speech) (HMM; Hidden Markov Model) (machine learning) deep learning (DNN; Deep Neural Network) 1:1 3 17. : (excitation) HMM TTS : 2012.08 ~ 2013.06 (11) (Time-Frequency Trajectory Excitation, TFTE) - / I (1) . HMM TTS . (TFTE) 18. : / : 3,000/20 (TFTE) ( ) : 20 X : I (2) . . . &lt; &gt; : TFTE : : PoN : STRAIGHT : 19. : (DNN; Deep Neural Network) TTS : 2013.11 ~ 2014.06 (8) DNN , , DNN II (1) . (DNN) . 20. DNN TTS DNN - - - DNN II (2) 21. : / : 3,000/20 DNN HMM II (3) TFTE-DNN (duration: known) TFTE-HMM 512*3 1024*3 (LSD) [dB] 3.10 3.12 5.27 SEW (NMSE) 0.31 0.31 0.39 F0 (RMSE) [Hz] 24.11 24.12 26.91 22. : (DNN) TTS : 2014.09 ~ 2015.06 (10) TTS ? ? TTS DNN ? III (1) , Paris la tour Eiffel. 23. TTS () (mapping rule) Deep learning : III (2) . DB DB DB DB DB DB DB 24. HCI (Human Computer Interface or Interaction) , IT , / , 25. Q&amp;A </p>