代码

进入：~/kaldi/egs/aishell/v1

# 地址变量（引号可省略，但是=之间不可有空格！）data=/export/a05/xna/datadata_url=/resources/33# 载入引用文件（使用.，类似import）. ./cmd.sh. ./path.sh# 设置bash出错则退出set -e # exit on error# 下载和解压数据（应该是两者选一）local/download_and_untar.sh $data $data_url data_aishelllocal/download_and_untar.sh $data $data_url resource_aishell#（download_and_untar.sh <数据/数据下载地址> <数据输出位置>）# 数据准备（应该是提取出wav数据，并且生成“说明文档”？）local/aishell_data_prep.sh $data/data_aishell/wav $data/data_aishell/transcript# 提取MFCC特征（建议使用一个大容量的磁盘来储存mfcc）# 每个wav首先被分为很多帧，每帧用一串数字来表示，mfcc就是这串数字mfccdir=mfccfor x in train test; dosteps/make_mfcc.sh --cmd "$train_cmd" --nj 10 data/$x exp/make_mfcc/$x $mfccdir# make_mfcc.sh <数据所在文件夹> <日志文件夹> <mfcc输出文件夹>；# -- 符号后面接的都是超参数设置sid/compute_vad_decision.sh --nj 10 --cmd "$train_cmd" data/$x exp/make_mfcc/$x $mfccdir# compute_vad_decision.sh <数据所在文件夹> <日志文件夹> <mfcc输出文件夹>utils/fix_data_dir.sh data/$x# fix_data_dir.sh 只需要一个参数done# 训练 diag ubmsid/train_diag_ubm.sh --nj 10 --cmd "$train_cmd" --num-threads 16 \data/train 1024 exp/diag_ubm_1024# 训练 full ubmsid/train_full_ubm.sh --nj 10 --cmd "$train_cmd" data/train \exp/diag_ubm_1024 exp/full_ubm_1024# 训练 ivector sid/train_ivector_extractor.sh --cmd "$train_cmd --mem 10G" \--num-iters 5 exp/full_ubm_1024/final.ubm data/train \exp/extractor_1024# 提取 ivectorsid/extract_ivectors.sh --cmd "$train_cmd" --nj 10 \exp/extractor_1024 data/train exp/ivector_train_1024# 训练 plda$train_cmd exp/ivector_train_1024/log/plda.log \ivector-compute-plda ark:data/train/spk2utt \'ark:ivector-normalize-length scp:exp/ivector_train_1024/ivector.scp ark:- |' \exp/ivector_train_1024/plda# --- ↓ split the test to enroll and eval ↓ ---mkdir -p data/test/enroll data/test/evalcp data/test/{spk2utt,feats.scp,vad.scp} data/test/enrollcp data/test/{spk2utt,feats.scp,vad.scp} data/test/evallocal/split_data_enroll_eval.py data/test/utt2spk data/test/enroll/utt2spk data/test/eval/utt2spk# split_data_enroll_eval.py # 将测试集（utt2spk）拆分为注册集和验证集# 输入：test utt2spk；输出：enroll utt2spk, eval utt2spk# 对每一个说话人，随机选取其下的3条音频作为验证集，剩下其它的用作验证集测试trials=data/test/aishell_speaker_ver.lstlocal/produce_trials.py data/test/eval/utt2spk $trials# local/produce_trials.py# 这个脚本用来生成 trails 文件：uttid spkid target|nontargetutils/fix_data_dir.sh data/test/enrollutils/fix_data_dir.sh data/test/eval# utils/fix_data_dir.sh# This script makes sure that only the segments present in# all of "feats.scp", "wav.scp" [if present], segments [if present]# text, and utt2spk are present in any of them.# It puts the original contents of data-dir into# data-dir/.backup# --- ↑ split the test to enroll and eval ↑ ---# 提取注册集的 ivectorsid/extract_ivectors.sh --cmd "$train_cmd" --nj 10 \exp/extractor_1024 data/test/enroll exp/ivector_enroll_1024# 提取验证集的 ivector sid/extract_ivectors.sh --cmd "$train_cmd" --nj 10 \exp/extractor_1024 data/test/eval exp/ivector_eval_1024# 计算 plda 分数$train_cmd exp/ivector_eval_1024/log/plda_score.log \ivector-plda-scoring --num-utts=ark:exp/ivector_enroll_1024/num_utts.ark \exp/ivector_train_1024/plda \ark:exp/ivector_enroll_1024/spk_ivector.ark \"ark:ivector-normalize-length scp:exp/ivector_eval_1024/ivector.scp ark:- |" \"cat '$trials' | awk '{print \\\$2, \\\$1}' |" exp/trials_out# ~kaldi/src/ivectorbin/ivector-plda-# 使用PLDA模型计算trails的对数似然比（og-likelihood ratios）# 输入（trials-file）："<key1> <key2>\n"# 输出： "<key1> <key2> [<dot-product>]\n"# 计算 EER awk '{print $3}' exp/trials_out | paste - $trials | awk '{print $1, $4}' | compute-eer -# Result# Scoring against data/test/aishell_speaker_ver.lst# Equal error rate is 0.140528%, at threshold -12.018exit 0

收获

linux shell 脚本的一些基础知识：比如$变量、<赋>=<值>、--超参数设置、.载入kaldi做说话人识别的大致框架

参考：

从说话人识别demo开始学习kaldi–（1）run.shkaldi入门：搭建第一个中文ASR (AISHELL-1)