HTK使用方法簡介(一)使用全字聲音單位(不含間隔音)李立民目標: 設計及製作連續數字語音辨認系統 進行語音辨認實驗 系統性能評估分析Step 1: Feature Extraction 將實驗所需的訓練音及測試音波形檔 (*.wav)轉換成特徵檔(*.mfcc) 建立一個設定檔fea_extract.cfg來指定轉換之 輸入與輸出格式。 建立一個輸入與輸出檔路徑表檔案 io_filelist.txt 執行HCopy -C fea_extract.cfg -S io_filelist.txt# Coding parameters SOURCEKIND = WAVEFORM SOURCEFORMAT = WAV SOURCERATE = 1000 #10KHz TARGETKIND = MFCC_E_A_D TARGETRATE = 100000.0 # (10ms) WINDOWSIZE = 250000.0 # (25ms) ZMEANSOURCE = F USEHAMMING = T PREEMCOEF = 0.97 NUMCHANS = 26 CEPLIFTER = 22 NUMCEPS = 12 ENORMALISE = F輸入及輸出格式設定檔 fea_extract.cfg輸入及輸出檔路徑表 io_filelist.txt輸入檔案路徑 輸出檔案路徑wavS500_00.wav mfccS500_00.mfcc wavS500_01.wav mfccS500_01.mfcc wavS500_02.wav mfccS500_02.mfcc wavS500_03.wav mfccS500_03.mfcc wavS500_04.wav mfccS500_04.mfcc wavS500_05.wav mfccS500_05.mfcc wavS500_06.wav mfccS500_06.mfcc wavS500_07.wav mfccS500_07.mfcc wavS500_08.wav mfccS500_08.mfcc wavS500_09.wav mfccS500_09.mfcc wavS500_10.wav mfccS500_10.mfcc 使用Hlist觀察特徵向量檔內容 執行 Hlist -h -o mfccS100_00.mfcc temp.txt 觀察temp.txt 內容- Source: mfccS100_00.mfcc -Sample Bytes: 156 Sample Kind: MFCC_E_D_A_KNum Comps: 39 Sample Period: 10000.0 usNum Samples: 57 File Format: HTK - Observation Structure - x: MFCC-1 MFCC-2 MFCC-3 MFCC-4 MFCC-5 MFCC-6 MFCC-7 MFCC-8 MFCC-9 MFCC-10MFCC-11 MFCC-12 E Del-1 Del-2 Del-3 Del-4 Del-5 Del-6 Del-7Del-8 Del-9 Del-10 Del-11 Del-12 DelE Acc-1 Acc-2 Acc-3 Acc-4Acc-5 Acc-6 Acc-7 Acc-8 Acc-9 Acc-10 Acc-11 Acc-12 AccE - Samples: 0-1 - 0: -5.628 4.174 1.699 -3.216 -4.589 -9.942 3.245 -7.514 4.315 -6.942-1.728 -4.513 10.109 0.224 0.128 1.119 0.065 0.770 0.565 -0.357-0.866 -0.081 -0.587 0.995 -2.256 0.511 0.764 0.742 0.243 0.194-0.059 -0.754 -0.393 -0.419 -0.640 0.093 0.025 0.343 0.367 1: -5.850 4.087 0.675 -4.400 -3.026 -11.016 1.422 -12.226 -1.280 -15.1146.278 -15.367 10.593 1.654 1.737 1.957 0.505 1.100 -0.416 -1.794-2.036 -1.792 -0.435 0.252 -1.784 1.479 0.864 0.852 0.112 0.372-0.293 -1.167 0.028 -0.225 -0.453 0.274 -0.151 1.024 0.394Step 2: 音檔標音並蒐集會用到的聲音單位d0 ling d1 i d2 er d3 san d4 su d5 wu d6 liou d7 chi d8 ba d9 jiou SILENCE sil建立標音字典 word_to_syllable.dic 本例標音字典中每個音都會用到 用HLEd程式將 word_trans.mlf 轉成 syllable_trans.mlf的標音檔 編輯轉換規則檔dig_to_syllable.led，內容 含EX及IS sil sil 執行HLEd -d word_to_syllable.dic -i syllable_trans.mlf dig_to_syllable.led word_trans.mlfword_to_syllable.dic#!MLF!# “mfcc/S5/00_00.lab“ sil ling sil . “mfcc/S5/00_01.lab“ sil i sil . “mfcc/S5/00_02.lab“ sil er sil .“mfcc/S5/00_10.lab“ sil liou su su liou i san sil . “mfcc/S5/00_11.lab“ sil liou ba i sil .syllable_trans.mlfStep 3: 建立HMM model的架構(產生初始HHM)模型架構放哪裡?造好存放初始HHM架構的目錄 mkdir model_struct del /q model_struct 系統共使用幾種架構？哪些聲音的HMM檔要使用同一種架構 為每個架構建一個 “架構設定檔”及 “模型表列檔” 注意:模型表列檔最後要有一個空白列例 model_struct1.pcf , hmmlist1.txt 設定各數字音模型(6個狀態)model_struct2.pcf , hmmlist2.txt 設定靜音模型(3個狀態) 架構設定檔內含 HMM狀態數, 連接結構, 要產生哪幾個聲音的HMM檔 (HMMList欄位) 產生的HMM檔要放在哪個目錄(OutDir欄位) 執行產生各種架構HMM檔之指令，例: Perl MakeProtoHMMSet model_struct1.pcfPerl MakeProtoHMMSet model_struct2.pcf 將各個模型表列檔合併成hmmlist.txt以便將來使用hmmlist1.txtCOMMENT This PCF produces a single mixture, single stream prototype systemhsKind: P covKind: D nStates: 6 nStreams: 1 sWidths: 39 mixes: 1 parmKind: MFCC_D_A_E vecSize: 39 outDir: model_struct hmmList: hmmlist1.txtmodel_struct1.pcfhmmlist2.txtsilling i er san su wu liou chi ba jiouo 39 1 39 h “ling“82 1 11 1.0000390.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 391.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 3 1 11 1.0000390.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 391.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ling的model structureStep 4: 計算training data的mean及 variance 作為模型初值 建立訓練音路徑列表檔filelist_odd.txt 造一個模型根目錄models, 及其下子目錄 hmm0hmm20以存放各階段的模型檔 mkdir models del /s /q models* for /L %m in (0,1,20) do mkdir modelshmm%m 執行下列指令以將各模型檔初值放在hmm0目 錄下 for /f %m in ( hmmList.txt ) do HCompV -f 0.01 -m -S filelist_odd.txt -M models/hmm0 model_struct/%mfilelist_odd.txt mfccS501_00.mfcc mfccS501_01.mfcc mfccS501_02.mfcc mfccS501_03.mfcc mfccS501_04.mfcc mfccS501_05.mf