kaldi 實(shí)時(shí)流 語(yǔ)音識(shí)別 語(yǔ)音評(píng)測(cè)交流QQ 1183214565

?訓(xùn)練步驟

### 1. train a monophone system

? ? steps/train_mono.sh --boost-silence 1.25 --nj 5 --cmd "$train_cmd" \

? ? data/train_500short data/lang_nosp exp/mono

### 2.train a first delta + delta-delta triphone system on all utterances

? ? steps/train_deltas.sh --boost-silence 1.25 --cmd "$train_cmd" \

? ? 2000 10000 data/train_clean_5 data/lang_nosp exp/mono_ali_train_clean_5 exp/tri1

### 3.train an LDA+MLLT system.

? ? steps/train_lda_mllt.sh --cmd "$train_cmd" \

? ? --splice-opts "--left-context=3 --right-context=3" 2500 15000 \

? ? data/train_clean_5 data/lang_nosp exp/tri1_ali_train_clean_5 exp/tri2b

### 4.Train tri3b, which is LDA+MLLT+SAT

? ? steps/train_sat.sh --cmd "$train_cmd" 2500 15000 \

? ? data/train_clean_5 data/lang_nosp exp/tri2b_ali_train_clean_5 exp/tri3b

### 5. Now we compute the pronunciation and silence probabilities from training data,and re-create the lang directory.

? ? steps/get_prons.sh --cmd "$train_cmd" \

? ? data/train_clean_5 data/lang_nosp exp/tri3b

### 6. Train a chain model

? ? local/chain/run_tdnn.sh --stage 0

### 實(shí)時(shí)流

? this->feature_pipeline = new OnlineNnet2FeaturePipeline(*(this->feature_info));

? this->decoder = new SingleUtteranceNnet3Decoder(*(this->decoder_opts), *(this->trans_model),

? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? *(this->decodable_info),

? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? *(this->decode_fst), this->feature_pipeline);