よēだ摸琌沮琘粂琿﹚ㄤ弧杠┮妮よē跋よ猭琌ㄆм砃璶舱Θ场だ{1}そ诀闽盎琩瘆﹟ゼ磝搐デ竜尔好ō薄ㄒ筿獺禕腇纞奶耑单ンいㄌ沮デ竜尔好よē﹚ㄤ┮妮膟砮Τ罽盎琩絛瞅玛﹚デ竜尔好ō よē笆だ摸よ猭耕咎竒喷肚参侩醚よ猭ㄣΤ祏弘蔼芠┦眏单纔墩ヘ玡ノ膀宾焊繵瞯眯╰计{2}絬┦稰箇代╰计{3}单┏糷羘厩疭┖粂贺醚家秈︽粂贺醚よēだ摸单闽╯镣Ч到ゅ膍{4}硄筁舱麓疭┖琈甮呼蹈self-organizing feature map, SOM癸宾焊眯╰计Mel-scale frequency cepstral coefficients, MFCC秈︽礚菏服籈摸ㄏノや秖诀support vector machines, SVMЧΘよē侩醚だ摸ゅ膍{5}玥膀秖丁家vector space model, VSM癸だ澄羘厩虫じ秈︽疭┖ボノ家癡絤程沧眔纔夹计沮栋だ摸狦狝祏よē醚弘拜肈ゅ膍{6}ノ瞏厩策よ猭眔獺┖耕瞺繴疭┖ㄓ竒ㄥ瞏厩策家砆約獂莱ノ粂だ摸ヴ叭いㄤ璶癪膍琌﹍疭┖矗ㄑ贺蔼糷疭┖ボ承甧跋だ疭┖丁粄竒筁瞏竒呼蹈矪瞶疭┖丁い璶笲ノ睱糷softmaxだ摸竟碞笷醚狦皐癸粂ㄣΤ疭┖妮┦ゅ膍{7}だㄏノ碻吏竒呼蹈recurrent neural network, RNN籔祏戳癘拘呼蹈long short-term memory, LSTM龟瞷膀跋だ迭よēだ摸ゅ膍{8}龟喷い祇瞷ㄏノ┏糷羘厩疭┖穦メア场だ疭┖獺ノ膀厩策よ猭PTN龟瞷粂贺だ摸眔纔LSTMだ摸弘ゅ膍{9}皐癸よēだ摸ヴ叭硄筁MatConvNetㄣ絚篶縩竒呼蹈convolutional neural net-works, CNNノ虫竊蝴羘眯瓜秈︽癡絤喷靡癸默よē龟瞷虏虫だ摸ヴ叭瓃ゅ膍╯Θ狦碝т続瞏厩策家琜篶膀狠狠よēだ摸瞏羘厩家ㄣΤ﹚︽┦籔╯基 ぃ呼蹈家羘厩獺┖琌紇臫よēだ摸莱ノ醚絋瞯璶ぇセゅ盢ヘ玡盽ノ呼蹈家挡篶SOMRNNLSTMCNNだノpython粂ēTensorflowキ秈︽龟瞷篶默ずよē畐ノ家癡絤喷靡籔代刚程秈︽ぃ家丁癡絤栋纔て呼蹈秸把だ摸┦癸ゑだ猂龟喷喷靡よē醚呼蹈家羘厩獺┖ゼㄓ秨祇砞よē笆﹚╰参矗ㄑ╯膀娄 1羘厩家盽ノ呼蹈挡篶 粂醚羘厩家烩办肚参竒呼蹈GMM-HMM家{10}瘤硄筁耎碫よΑ莉眔疭┖獺耎碫Τぃ厩策瞏糷獶絬┦疭┖繦帝瞏厩策呼蹈砍癬厩Hinton盢瞏竒呼蹈DNN-HMM莱ノ粂厩烩办{11}莉眔醚┦陪帝矗ど钡帝箇代粂计沮吏硈钡RNN砆矗Θ笲ノ羘厩家{12}赣家パㄌ苦癘拘拜肈玻ネ辫ア辫脄拜肈膀厩Gonzalez单{13}矗ミRNN跑砰LSTM醚よēㄏ醚┦矗ど15%τ粂醚烩办LSTM-HMMDNN-HMM家Θ程約獂莱ノ挡篶CNN砆Θ莱ノ粂醚籔祏粂贺醚{14-15}CNNぃэ秈家VGG-NetGoogleNet㎝ResNet矗常CNNよē醚矗ㄑ╯よ 1.1 RNN RNN琌贺¨ず场患耴〃竒呼蹈瓜1┮ボよē獺┖筁祘いDNN惠璶硄筁耎碫よΑㄓ笷粂ゅ獺怠琌㏕﹚τRNN盢碫徊籔セ碫獺癬呼蹈い厩策 粂獺瑈笆筁祘いぃ﹚碻吏絋玂呼蹈Τ匡拒纗獺ず场篈パRNN呼蹈–˙穝镑癸丁粂琿秈︽家惠璶猔種琌竊翴计秖糤矗ど家笷琌穦盿ㄓ筁览の璸衡瞯单翴匡拒続竊翴计琌羘厩家闽龄翴ぇ 1.2 LSTM LSTM衡猭秆∕RNN衡猭呼蹈癡絤筁祘いΜ滥拜肈旧璓辫ア㎝辫脄糤篈筁耾瓜2┮ボЧ俱LSTM挡篶瞶秆眖オ獺秖瑈珹丁˙t矪块秖Xt丁˙t-1矪ぇ玡留旅篈秖ht-1㎝ぇ玡癘拘虫じ篈丁˙t-1矪絪絏秖Ct-1籔肚参RNNゑLSTM秆舰留旅篈ht㎝纗虫じCtㄏず甧秖陆す砛呼蹈硄筁承狡馒ずㄓ厩策籔矪瞶丁丁筳ず块獺 1.3 CNN 繵繵眯瓜挡篶疭┖镑▆┖羘初春计沮瞏厩策家块CNN癬方糷稰诀呼蹈multilayer perceptron, MLPノ縩笲衡㎝て稱硄筁祡波硈钡舦ㄉ㎝Τ匡拒┾妓矪瞶酚块糷縩糷て糷硈硄糷块糷抖Θ家琜篶瓜3┮ボ ノ縩丁丁キ簿ぃ跑┦狝よē獺腹妓┦琌︽礛τヘ玡よēだ摸╯烩办CNN琜篶膀セ衡猭耕ㄌ苦计沮栋癡絤秖 2龟喷 龟喷篶ノ癡絤厩策默よē妓セ畐钡帝砞璸龟喷よミ┦蝶基夹程硄筁ぃ厩策呼蹈家把计砞竚秨甶┦蝶基备ㄣΤよē疭┖粂计獺腹挡篶疭翴ミ膀瞏厩策羘厩家よ猭 2.1よē代刚栋 よē畐篶ノ眖Τ戈方い备才癡絤璶―粂粂计沮栋よ篶默ずよē计沮栋默よē跋璓だ瞐﹛杠粂籔い﹛杠よē跋ㄤいい﹛杠だガ跋办璓畗盝綞场籔硈冻翠场粂だガ跋办璓默奎盽跋玭硄玭场箃马う锭瞐﹛杠玥璓だガ默ㄤ郡カ跋粂灿だび打㎝默烈跋ず璶パび打瑂钞籔默涵古舱Θ瞐﹛杠灿だ籔瑇盻ヘ玡Τ皐癸瓣よēだ跋τ砞璸そ秨计沮栋龟筋筁祘い祇瞷场だ疭﹚跋ご笷ぃ秨甶瞏厩策癡絤┮惠计沮秖计沮栋夹帽帝ぃЧ俱灿だぃì单拜肈膀龟喷い砞璸パ跋┦玌粂舱Θゅセ癸瓃よē跋ず105弧杠酚セ跋玌粂ゅセ瓃粂秈︽魁栋盢いゅよē粂醚计沮畐King-ASR-M-004よē默秏眣喘跌繵戈干粂┮Τ粂妓セА竒筁靖把计砏俱繰纐琿ち埃箇矪瞶ΑwavΑ妓瞯11025 Hz妓计16 bit10 s璸眔默よē膀娄粂妓セ6036ㄒ沮弧杠┮妮跋癸–ㄒ粂妓セ睰耴妮5摸夹帽 瓜菠 瓜1RNN呼蹈家挡篶 Fig.1 RNN network model 瓜菠 瓜2LSTM呼蹈家挡篶 Fig.2 LSTM network model 瓜菠 瓜3CNN呼蹈家琜篶 Fig.3 CNN network model 1舱Θ栋ゅセ跋场だㄥ玌粂ㄤい玌粂ノ簙粂猔よ猭绢Α 龟喷い酚6?2?2ゑㄒちだ癡絤秨祇籔代刚栋癸よē代刚栋秈︽秈˙絪胯だ摸硄筁芭胯Θ3 s籔10 s代刚栋癸代刚计沮栋いだフ靖羘Θ獺靖ゑだ310 dB代刚栋眖τ程沧Θ4代刚计沮栋–代刚栋いА1207代刚妓セ 1默跋场だㄥ玌粂 Table 1 Typical dialects spoken in Jiangsu province | よē跋 | 簙粂猔 | 炊硄杠 | | 瑂钞
| ㄕㄚㄅㄢ | 糉甡 | | ㄑㄧㄝㄎㄡ | 混杠 | | ㄐ丨ㄝㄑ丨ㄤ | 郴糒 | | ㄏㄢㄧㄉㄨㄣ | 逗 | | ㄎㄨㄣㄍㄠ | 何谋 | | ㄤㄙㄢㄏㄨㄛ | 短矹 | | 瑇盻
| ㄑㄩㄒㄧㄢ | 矻癈 | | ㄦㄧㄗㄧ | 跑篈 | | ㄒㄧㄚㄎㄜㄗㄧ | ぺ | | ㄉㄧㄥㄅㄢㄦ | Ь | | ㄎㄢㄎㄨㄥ | 弧晾 | | ㄈㄢㄒㄧㄢ | 癚菇 | | 畗瞐
| ㄍㄟㄒㄧㄝ | | | ㄘㄞㄏㄨㄞ | を痜 | | ㄌㄧㄠㄘㄞ | тㄆ | | ㄉㄤㄧㄤ | い丁 | | ㄌㄚㄍㄨㄚ | 册ぱ | | ㄌㄞㄉㄞㄏㄨㄥ | 缅唤 | | 默涵古
| ㄒㄧㄤㄇㄚ | 琜 | | ㄒㄧㄠㄋㄧㄤㄩ | | | ㄖㄨㄢㄑㄧ | ぃ瞡 | | ㄧㄜㄎㄨㄞㄊㄡ | 衬边 | | ㄌㄞㄇㄠ | 繷祇 | | ㄙㄤㄍㄚㄌㄤ | 硙刁 | |
| ㄧㄚㄌㄧ | 边 | | ㄨㄒㄩ | 獽 | | ㄍㄚㄍㄚ | 產 | | ㄕㄨㄟㄐㄧ | 獵奠 | | ㄖㄡㄅㄞ | | | ㄒㄧㄢㄉㄡㄗㄞ | |
2.2蝶基夹 ㄏノミ默よē粂栋籔篶呼蹈家秈︽癡絤喷靡籔代刚パミ呼蹈家┮秆∕超栋だ摸拜肈絋瞯ぃ琌斑蝶基夹σ納だ摸妓セ计癸А颗セ龟喷ノだ摸蝶基夹PRFぃ盿舦ЩキАMacro-F1㎝稬キАMicro-F1龟喷家蝶基夹{16} 赣蝶基夹盢だ摸拜肈だ秆Nだ摸拜肈玻ネ睼瞔痻皚–睼瞔痻皚い琘跋Mよē妓セ砆タ絋だ摸M癘痷タㄒTP ぃ妮琘跋Mよē妓セ砆岿粇だ摸M癘安タㄒFP妮跋よēM妓セ砆岿粇だ摸獶M跋よē癘安璽ㄒFN 1Micro-F1夹 Micro-F1琌膀稬琩瞯㎝稬琩瞯秸㎝キА﹚竡璸衡そΑ 稬琩瞯そΑ菠1 稬琩瞯そΑ菠2 micro-F1=2⊙microP⊙microR/microP+microR3 ㄤいそΑ菠そΑ菠そΑ菠ボ睼瞔痻皚い癸莱计衡计キА 2Macro-F1夹 惠璶沮–だ摸睼瞔痻皚璸衡ㄤ琩瞯P㎝琩瞯R礛璸衡キАぃσ納妓セ舦莉眔macroP籔macroR程―眔Macro-F1ㄣ砰そΑ Щ琩瞯そΑ菠4 Щ琩瞯そΑ菠5 micro-F1=2⊙microP⊙microR/microP+microR6 ㄤいn妓セだ摸计セ龟喷い沮よē跋计n=5ЩキАMacro-F1籔稬キАMicro-F1计は琈家铆﹚┦ㄤ絛瞅[0,1]だσ納家獂て玡矗ㄤ计禫家┦禫铆﹚ 2.3呼蹈家龟瞷 盢膀SOMRNNLSTMCNN4摸家琜篶沽刚程纔把计砞竚よ秈︽癡絤籔喷靡 2.3.1计沮砏俱 皐癸SOMRNNLSTM家ノkaldiㄣ矗MFCC疭┖匡拒ㄏノMFCCのㄤ顶畉だ顶畉だ╰计36蝴疭┖呼蹈块ノ撑腑跑传碫25 ms碫簿10 ms怠摸簙诂怠ㄏ块疭┖骸ìCNN家蝴计沮Α龟喷いㄏノpythonいPIL瓜钩矪瞶畐ㄧ计癸瓜秈︽砏俱ㄏノspecgramㄧ计よ猭盢蝴繵计沮锣传Θ蝴眯瓜Α瓜4钡帝ㄏノcropㄧ计よ猭埃丁蝴いフ跋办程盢繵眯瓜い繵瞯蝴籔丁蝴玂璓ㄏノresizeㄧ计よ猭砏俱瓜へ224⊙224 瓜菠 瓜4疭┖繵眯瓜 Fig.4 Characteristic frequency spectrum 2.3.2呼蹈砞竚 1SOM家 龟喷硄筁SOM癸36蝴MFCC疭┖秈︽籈摸匡拒挡篶繧程て玥や秖诀SVMだ摸竟パSOM-SVM家ぃ妓セ疭┖蝴紇臫妓セ丁兵ン眔だ摸狦 竒筁龟喷だ猂SVMだ摸竟ㄧ计匡拒匡拒蔼吹畖膀ㄧ计ㄧ计胓籃╰计C=150癡絤粇畉=0.002 2RNN\LSTM家 癡絤い祇瞷厩策硉瞯留旅糷单禬把计砞竚癸家だ摸┦穦玻ネ耕紇臫ぃ耞沽刚匡拒留旅糷糷计3絋瞯程ㄎ篈龟喷い沽刚ぃ耞糤留旅糷竊翴计祇瞷讽竒じ计ヘ笷128家癡絤丁陪帝糤竒じ计ヘ程沧匡拒砞竚256匡拒sigmoid縀ㄧ计ノユ鎖穕アㄧ计キ颗Μ滥硉籔Ω计闽╰砞﹚batch_size64ЧΘΩepoch惠60Ω厩策瞯癐搭匡ノ㏕﹚˙癐搭郸菠硄筁ぃ耞秸刚砞﹚厩策硉瞯﹍0.05–20Ωepoch0.01纔て竟玥﹚竡AdamOptimizer 瓃把计兵ン穕アㄧ计癡絤15Ωepochオ镣キ铆龟喷莱RNN/LSTM挡篶膀娄硄筁膙块家Α癸竊翴┮妮よē跋办秈︽だ摸 3CNN家 把σゅ膍{17}縩糷挡篶ミ龟喷家赣家パ6縩糷籔3硈钡糷舱Θ度材1縩糷糤1て糷匡拒maxout縀ㄧ计CNN家块224⊙224⊙3匡拒縩へ3⊙3˙1てへ匡拒3⊙1度繵办絛瞅秈︽て padding把计砞﹚sameЧΘ–Ω縩ㄏノ箂恶zero-padding北疭┖眯瓜家把计秖4.1M竒筁は狡龟喷絋粄禬把计砞﹚mini-batch=32Ω计epoch=150ㄤい玡100Ω厩策硉0.0550Ω厩策硉瞯0.005程硈钡糷块硄筁softmax璸衡篈喷阀瞯眔だ摸挡狦 3挡狦籔癚阶 喷靡ぃ计沮栋妓セ计秖獺靖ゑ籔妓セ癸よēだ摸玻ネ紇臫匡拒4摸竒呼蹈家癸4摸代刚计沮栋秈︽默ずよēだ摸代刚ㄏノsklearn畐ㄧ计f1_score璸衡眔ぃ计沮栋よēだ摸睼瞔痻皚籔蝶基夹Micro-F1Macro-F110 s/10 db计沮栋兵ンLSTM家籔CNN家だ摸睼瞔痻皚瓜5┮ボ酚そΑ1~6膀眔睼瞔痻皚秈˙璸衡眔ぃ代刚栋呼蹈家蝶基夹Micro-F1Macro-F1ЩキАMacro-F1计甧代刚栋いよē妓セ计ぃА颗紇臫ㄣ砰2┮ボ 瓜菠 瓜510s/10db计沮栋よēだ摸睼瞔痻皚a. LSTM; b. CNN Fig.5 Confusion matrices with 10s/10dB dataset for dialect identification a. LSTM; b. CNN 2ぃ代刚栋呼蹈家┦蝶基 Table 2 Performance evaluation of network models with different test aggregates | 家
| Micro-F1 | Macro-F1 | | 3db | 10db | 3db | 10db | | 3s | 10s | 3s | 10s | 3s | 10s | 3s | 10s | | SOM | 0.522 | 0.565 | 0.701 | 0.706 | 0.513 | 0.568 | 0.695 | 0.701 | | RNN | 0.659 | 0.672 | 0.717 | 0.723 | 0.647 | 0.667 | 0.718 | 0.720 | | LSTM | 0.747 | 0.780 | 0.885 | 0.931 | 0.711 | 0.769 | 0.876 | 0.927 | | CNN | 0.824 | 0.831 | 0.849 | 0.876 | 0.819 | 0.821 | 0.842 | 0.860 |
眖2龟喷挡狦いLSTM家莉眔程ㄎ蝶基夹LSTM籔CNN家蝶基夹陪纔SOM籔RNN家SOM竒呼蹈镑耕続莱セ龟喷计沮栋兵ンだ摸SOM竒呼蹈陪癸妓セ獺靖ゑ庇稰3db代刚栋醚絋┦耕畉RNNよē醚絋瞯眔﹚矗ど皐癸粂い丁丁筳耕璶獺┖ぃìパLSTMノō家挡篶虫じ疭翴狝RNN瓃翴10s代刚栋眔程ㄎ蝶基夹 CNN家醚絋瞯耕铆﹚ぃ粂琿紇臫癸粂靖羘稦耑ㄣΤ耕眏緗次┦硂琌パCNN家い縩て单獶絬┦跑传巨ㄣΤ獶絬┦笷癸よēだ摸莱ノ獺┖ㄣΤ耕览惠璶猔種琌CNN醚硉ぃ骸ì龟┦璶― 龟喷挡狦陪ボк靖┦龟喷┮ノ呼蹈家ぃ獺靖ゑ计沮栋兵ン眔蝶基夹Micro-F1Macro-F1籔獺靖ゑ计А瞷タ闽计沮妓セ籔蝶基夹タ闽闽╰硂呼蹈家祏粂獺矗矗どк靖緗次┦眔﹚秈˙粂妓セ籔獺靖ゑご礛琌紇臫家だ摸弘璶ぇ 4挡阶籔甶辨 ノ瞏厩策琜秆∕よē笆醚ヴ叭琌ゼㄓ璶╯よセゅ╯SOMRNN LSTMCNN呼蹈家羘厩家龟瞷籔獺┖だ猂硄筁龟喷挡狦だ猂ゑ耕阶靡CNN LSTM呼蹈家АㄣΤよē羘厩獺▆┖CNNк靖耕眏甶ボCNN籔LSTM羛莱ノゼㄓ秈︽よē笆だ摸Ω秨祇ㄣΤ﹚︽┦龟喷挡狦处ボ夹羘畐砞惠ミ砏絛羘妓セ栋夹羘醚м砃そ龟驹い祇揣驹矮矗ㄑ璶玂靡
Dear visitor,you are attempting to view a subscription-based section of lawinfochina.com. If you are already a subscriber, please login to enjoy access to our databases . If you are not a subscriber, please subscribe . Should you have any questions, please contact us at:
+86 (10) 8268-9699 or +86 (10) 8266-8266 (ext. 153)
Mobile: +86 133-1157-0713
Fax: +86 (10) 8266-8268
database@chinalawinfo.com
|