1. stage 1
在此之前,已经下载了130G的AMI数据在目录$AMI_DIR(wav_db/amicorpus)了。
# Prepare ihm data directories,if [ $stage -le 1 ]; thenlocal/ami_ihm_data_prep.sh $AMI_DIRlocal/ami_ihm_scoring_data_prep.sh $AMI_DIR devlocal/ami_ihm_scoring_data_prep.sh $AMI_DIR evalfi
ami_ihm_data_pre.sh脚本内部流程如下:
1) 在data/local/ihm/train文件夹下准备各种文件,如wav.scp;
2) 拷贝spk2utt utt2spk wav.scp text segments reco2file_and_channel到data/ihm/train文件夹中。
3) 检查文件格式:
utils/validate_data_dir.sh --no-feats data/ihm/train || exit 1;
ami_ihm_scoring_data_prep.sh内部流程和ami_ihm_data_pre.sh类似,只是处理的是dev和eval。
2. stage 2
# Feature extraction,if [ $stage -le 2 ]; thenfor dset in train dev eval; dosteps/make_mfcc.sh --nj 15 --cmd "$train_cmd" data/$mic/$dset data/$mic/$dset/log data/$mic /$dset/datasteps/compute_cmvn_stats.sh data/$mic/$dset data/$mic/$dset/log data/$mic/$dset/datadonefor dset in train eval dev; do utils/fix_data_dir.sh data/$mic/$dset; donefi
1)make_mfcc.sh的配置在文件conf/mfcc.conf里。
2)compute_cmvn_stats.sh: "Compute cepstral mean and variance statistics per speaker",
其output文件为cmvn_train.ark和cmvn.scp。
3. stage 3
if [ $stage -le 3 ]; then# Taking a subset, now unused, can be handy for quick experiments,# Full set 77h, reduced set 10.8h,utils/subset_data_dir.sh data/$mic/train 15000 data/$mic/train_15kfi