Initial HMM Model Feature cut lexicon 2239 HMM

  • Slides: 3
Download presentation
Initial HMM Model

Initial HMM Model

語音訊號 Feature , (查cut) 表示成所有中文語 言單元 經過lexicon字彙庫找出所 有的子音(22)、母音(39)、 左至右HMM 1. 412 0 0 29

語音訊號 Feature , (查cut) 表示成所有中文語 言單元 經過lexicon字彙庫找出所 有的子音(22)、母音(39)、 左至右HMM 1. 412 0 0 29 2. 313 0 30 31 3. 313 1 32 35 4. 313 2 36 40 5. 313 3 41 42 6. 313 4 43 59 7. 313 5 60 63 8. 313 6 64 66 9. 313 7 67 71 10. 412 0 72 79 11. 12. initial final 32 mixture VQ 32 mixture 32個µ、v VQ MLVQ u、v、c、det 32個µ、v MLVQ u、v、c、det

411. c begin Allocate initial & final model parameters m m m> I_MODEL 1

411. c begin Allocate initial & final model parameters m m m> I_MODEL 1 • transition prob • mean • variance • weight • determinant 1 m> F_MODEL 0 0 s s 1 1 s> F_STATE s> I_STATE 0 0 Open the database list file (fp_list) Read frames from the cls file Load lexicon VQ process • Remove model’s temp file • initialize Initialize MLVQ Generate new temp files • get mixture means model/state’s count • all speakers and all utterances • read feature and get delta term • SBR • read cut information • all syllables • all states • write state frames to the Read frames from the cls file • determine • variance • weight MLVQ process • get mean, variance, weight, • get mean, variance weight, determinant cls file of the corresponding state model • update model’s and state’s count Assign the mixture number of initial and final states Estimate the state transition prob Save 411 parameters of HMM model Delete old cls files end