Kaldi语音识别Lecture3.pdf

Speech REcognition - a practical guide Lecture 3 Phonetic Context Dependency Steps covered in this lecture $ cd ~/kaldi-trunk/rm/s3/ $ $ # Get alignments from monophone system. $ steps/align_deltas.sh data/train data/lang exp/mono exp/mono_ali $ $ # train tri1 [first triphone pass] $ steps/train_deltas.sh data/train data/lang exp/mono_ali exp/tri1 $ Aligning data with monophone system Training triphone system Weakness of “monophone” model Phones “sound different” in different contexts. Most strongly affected by phones immediately before/after. Simplest model of context dependency is to build separate model per “triphone” context. For 38 phones, #models required is 38x38x38 Too many models to train! Traditional context- dependency tree Build a “decision tree” for each “monophone” This follows the “Clustering and Regression Tree” (CART) framework Involves a “greedy” (locally optimal) splitting algorithm. Ask questions like “Is the left phone a vowel?” Is the right phone the phone “sh”? Models (HMMs) would correspond to the leaves “m” Left=vowel? Yes No Yes No Right=fricative? Square boxes correspond to Hidden Markov Models Traditional tree-building Train a monophone system (or use previously built triphone system) to get time alignments for data. For each seen triphone, accumulate sufficient statistics to train a single Gaussian per HMM state Suff. stats for Gaussian are (count, sum, sum- squared). Total stats size (for 39-dim feats): But not all ! triphones seen (38x38x38