Slide 1

Slide 1 text

Engineering Ի੠ೝࣝʹ͓͚ΔαʔόαΠυ։ൃ Shuta Ichimura Clova Developer Team @ Kyoto, LINE Corp.

Slide 2

Slide 2 text

Engineering • ࢢଜɹऩଠɹʢΠνϜϥɹγϡ΢λʣ • Clova։ൃνʔϜˏژ౎ • Ի੠ೝࣝΤϯδχΞ Since Sep. 2018 Ի੠ೝࣝཁૉٕज़։ൃ୲౰ Decoder։ൃ ԻڹϞσϧ։ൃ • झຯɿਆࣾ hopping (ژ౎ɺಸྑɺ۝भɺग़Ӣ etc..) SELF-INTRO

Slide 3

Slide 3 text

Engineering Server SCOPE OF TODAY NSpeech (Decoder) NLU Models (AM,LM) NVoice (Speech Synthesis) Clova Developer Team @ Kyoto Today’s scope is the ASR

Slide 4

Slide 4 text

Engineering CONTENTSɹ 1. Overview of the ASR 2. Developing Models(AM, LM) 3. Developing Decoder 4. Q & A

Slide 5

Slide 5 text

Engineering OVERVIEW OF THE ASR p(W|X) ≈ argmax p(X|W) p(W) AM Recog. Result LM Decoding Xfeatures = x1 , x2 . . xk Wwords = w1 , w2 . . wn Building in advance Decoding online

Slide 6

Slide 6 text

Engineering Extract Feature Features to Phone Word to Sentence Features (FBank, MFCC ,etc.) Phone Seq. to Word Phone Sequence Word Speech: FLOW OF THE ASR PROCESSɹ Text:ࠓ೔͸ӍͰ͢ AM (DNN) Lexicon (L.fst) HMM (HC.fst) Grammar (G.fst) LM (HCLG.fst) AM (*.nnet) use use use Training AM on Hadoop and GPU Training LM on Hadoop Developing Decoder Developing Server Side In ASR

Slide 7

Slide 7 text

Engineering DEVELOPING MODELS ։ൃεύϯ Acoustics Model(AM) جຊతʹఆظతͳϞσϧֶश͸ߦΘͳ͍ ԻڹతͳཁҼʢϚΠΫɺഎܠࡶԻetcʣ͕มΘΕ͹࠶ֶशΛߦ͏ Language Model(LM) جຊతʹຖि࡞੒͠Ξοϓσʔτ͍ͯ͠Δ ৽ޠͳͲੈͷதͷྲྀߦʹରԠ͠ͳ͍ͱ͍͚ͳ͍ͨΊ

Slide 8

Slide 8 text

Engineering Developing An Acoustic Model

Slide 9

Slide 9 text

Engineering On Hadoop Training Mono-Phone Training Tri-Phone Force Alignment Feat. And Transcript On GPUs Feat. And Tri-Phone Training Neural Net.(NN) NN based AM Pre-Process MapReduce Training On GPUs Tens of Millions ML Infer An Alignment Tri-Phone Feat. G2P etc.. Data A Couple of Days A Couple of Weeks Developing AM a m e sil-a+m a-m+e e+sil

Slide 10

Slide 10 text

Engineering Developing A Language Model

Slide 11

Slide 11 text

Engineering On Hadoop Counting Words Building N-Gram Transcripts(Corpuses) On CPU(Can’t Use Hadoop) Lex And N-Gram Building WFST WFST based LM Pre-Process Tens of Millions Lots of Processes Fixed prob. WFST Lex And N-Gram. Filtering,G2P etc.. A Day A day and Using A Couple of Hundred GB Memory Not suitable for distributed processing cuz of graph structure Developing LM

Slide 12

Slide 12 text

Engineering Developing Decoder

Slide 13

Slide 13 text

Engineering ௿ϨΠςϯγ͕ٻΊΒΕΔʢTAT:਺ඦmsecҎ಺ʣ ೝࣝύϥϝʔλνϡʔχϯάʢ଎౓ͱੑೳͷτϨʔυΦϑʣ ੑೳվળʢجຊతʹϞσϧ࠶ֶशͰ͸෭࡞༻͕େ͖͘ͳΓͦ͏ͳ՝୊ʹରͯ͠ରԠʣ υϝΠϯରԠ υϝΠϯϞσϧΛϕʔεʢେޠኮʣʹରͯ͠Ϛʔδ ˠઐ໳༻ޠͳͲͷҰൠతͰͳ͍୯ޠ͕ೝࣝͰ͖ΔΑ͏ʹ Corner-CaseରԠ ݱঢ়ͷAMɺLMͰؒҧ͑΍͍͢ൃ࿩ʹରͯ͠ॲஔΛߦ͏ ˠ AM scoreʹΑΔɺreranking ΍ pruning e.g.) ؒҧ͑΍͍͢έʔε 24 -> 24ճ ൃ࿩ͯ͠ͳ͍͕ɺʮճʯ͕ग़ྗ͞ΕΔ ͶʔͣʔΈʔ ˠ ϝʔϧ12 ؒԆͼͨ͠ൃ࿩ελΠϧͰσλϥϝͳ݁ՌͱͳΔ Text Normalization ೫໦ࡔϑΥʔςΟʔΤΠτ ˠ ೫໦ࡔ46 Head line new ˠ ϔουϥΠϯχϡʔε ޡೝࣝ෼ੳ ԻڹతͳཁҼ͔ݴޠతͳཁҼ͔Λ෼ੳ͢Δҝͷπʔϧ։ൃ Ϟσϧֶश͓ΑͼDecoder։ൃ΁ϑΟʔυόοΫ

Slide 14

Slide 14 text

Engineering LANGUAGE ETC.. • C/C++, python, scala etc.. • Kaldi, Hadoop

Slide 15

Slide 15 text

Engineering Thank you for your attention Questions ? END OF DOCUMENT Developers