Kaldi中每个脚本的简单解释

Kaldi中每个脚本的简单解释

utils/subset_data_dir.sh
分割数据,用于建立初始小模型,而后一步一步扩充

steps/train_mono.sh
单音素模型训练

steps/align.sh, steps/align_si.sh, steps/align_fmllr.sh
强制对齐

steps/train_sat.sh
说话人自适应,一般之后跟fmllr,第一个sat前用si或者fmllr,sat一般用两轮

steps/get_prons.sh
从训练数据中计算发音和静音概率,并重新创建lang目录,样例参见fisher_swbd/s5

steps/make_plp_pitch.sh
提取plp和pitch特征

steps/make_plp.sh
提取plp特征

utils/fix_data_dir.sh
数据规整

steps/make_fbank.sh
提取fbank特征,一般与local/nnet/run_dnn_fbank.sh组合使用

steps/make_mfcc.sh
提取MFCC特征,相较于fbank有损失

steps/compute_cmvn_stats.sh
cmvn,提取倒谱特征,语音识别时用

local/train_irstlm.sh
建lm的一个工具包

local/nnet3/xvector/prepare_feats.sh
cmvn,倒谱归一化,声纹识别时用

steps/align_fmllr.sh
fmllr对齐

steps/train_mmi.sh
句错误率最小化训练

steps/train_mpe.sh
字错误率(最小颗粒度)去训练

sid/train_diag_ubm.sh, sid/train_full_ubm.sh, steps/train_ubm.sh
ubm训练

steps/train_sgmm2.sh,steps/align_sgmm2.sh,steps/make_denlats_sgmm2.sh
sgmm训练

sid/compute_vad_decision_gmm.sh
基于VAD输出计算能量

sid/compute_vad_decision.sh
利用能量提取有效音频段

local/run_lmrescore.sh
利用RNN对LM重新打分

local/run_wpe.sh, local/run_beamformit.sh
麦克风阵列相关处理,用于数据增强,代码在chime5/s5b/run.sh中。此外,run.sh中还有加噪,混响相关代码

steps/data/reverberate_data_dir.py, steps/data/augment_data_dir.py
加噪,加混响相关操作,用于数据增强

chime6/s5_track2/local/train_diarizer.sh
训练xvector dnn

local/vtln.sh
用于消除男女声道长度差异

local/chain/run_tdnnf.sh,local/chain/run_tdnn.sh
tdnn训练脚本,tdnnf比tdnn两层中间多了层维数较低的中间层

local/nnet3/run_tdnn.sh
nnet3 TDNN

local/chain/run_tdnn_1g.sh
与tdnn_1f类似,但做了一些调整,样例在fisher_callhome_spanish/s5中

steps/train_deltas.sh
一般在tri1,也会在tri2,tri3

steps/train_lda_mllt.sh
LDA+MLLT,一般在tri2,tri3,tri2b,tri3b,看个人喜好命名

steps/train_quick.sh
在现有特征的基础上训练模型(不进行任何类型的特征空间学习)

local/run_sgmm2.sh
SGMM训练

local/nnet/run_dnn.sh D
NN训练

local/online/run_nnet2_ms.sh

local/csj_run_rnnlm.sh
日语重打分RNNLM训练

diarization/vad_to_segments.sh
音频做vad

diarization/score_plda.sh, diarization/cluster.sh
plda打分,根据打分分类,合并重复说话人。一般说话人id不明确的时候用

local/nnet3/xvector/prepare_feats_for_egs.sh, local/nnet3/xvector/run_xvector.sh, sid/nnet3/xvector/extract_xvectors.sh
CMVN,提取xvector特征

ivector-mean, ivector-compute-lda, ivector-compute-plda
lda和plda训练

ivector-plda-scoring
plda打分

sid/train_diag_ubm.sh, sid/train_full_ubm.sh, sid/train_ivector_extractor.sh
一般提取ivector,例子可见fame/v1

sid/init_full_ubm_from_dnn.sh, sid/train_ivector_extractor_dnn.sh, sid/extract_ivectors_dnn.sh
用dnn提取相关ivector特征,例子可见fame/v2

copy-feats
查看ark文件,一般文件合并时用

    分享到:

Leave a Reply

Your email address will not be published. Required fields are marked *