espnet学习(1):AN4数据集语音识别

最新推荐文章于 2025-05-16 09:49:25 发布

大雪001

最新推荐文章于 2025-05-16 09:49:25 发布

阅读量3k

点赞数 4

分类专栏： Espnet 文章标签：语音识别 espnet python 深度学习

本文链接：https://blog.youkuaiyun.com/LeavingBook/article/details/111512311

版权

step -1. 数据下载

# data
datadir=./downloads

if [ ${
   stage} -le -1 ] && [ ${
   stop_stage} -ge -1 ]; then
    echo "stage -1: Data Download"
    mkdir -p ${
   datadir}
    local/download_and_untar.sh ${
   datadir} ${
   data_url}
fi

通过运行local下的download_and_untar.sh脚本下载an4数据集，当然也可以自己从官网下载自己解压缩，下载后的数据集存放在downloads文件夹下
在这里插入图片描述

step 0~2. Kaldi格式数据准备

step0. 数据准备

# general configuration
dumpdir=dump   # directory to dump full features

# feature configuration
do_delta=false

# data
datadir=./downloads
an4_root=${
   datadir}/an4

train_set="train_nodev"
train_dev="train_dev"


if [ ${
   stage} -le 0 ] && [ ${
   stop_stage} -ge 0 ]; then
    ### Task dependent. You have to make data the following preparation part by yourself.
    ### But you can utilize Kaldi recipes in most cases
    echo "stage 0: Data preparation"
    mkdir -p data/{
   train,test} exp #递归创建data/train、data/test、exp目录

    if [ ! -f ${
   an4_root}/README ]; then #找不到an4数据集的情况
        echo Cannot find an4 root! Exiting...
        exit 1
    fi

    python3 local/data_prep.py ${
   an4_root} sph2pipe#运行local文件夹下的data_prep.py

	#生成train和test的wav.scp、text、utt2spk、spk2utt
    for x in test train; do
        for f in text wav.scp utt2spk; do
            sort data/${
   x}/${
   f} -o data/${
   x}/${
   f}
        done
        utils/utt2spk_to_spk2utt.pl data/${
   x}/utt2spk > data/${
   x}/spk2utt
    done
fi

#递归创建文件夹dump/train_nodev/daltafalse、dump/train_dev/daltafalse
feat_tr_dir=${
   dumpdir}/${
   train_set}/delta${
   do_delta}; mkdir -p ${
   feat_tr_dir}
feat_dt_dir=${
   dumpdir}/${
   train_dev}/delta${
   do_delta}; mkdir -p ${
   feat_dt_dir}

生成最重要的四个文件wav.scp、text、utt2spk、spk2utt
wav.scp文件保存了语音编号和该语音在系统中的绝对路径的位置
test文件保存了语音编号与该语音对应的转录文本
utt2spk文件保存了语音编号和说话人编号
spk2utt文件保存了说话人编号和该说话人所有语音编号的对应关系

step1. 特征提取

if [ ${
   stage} -le 1 ] && [ ${
   stop_stage} -ge 1 ]; then
    ### Task dependent. You have to design training and dev sets by yourself.
    ### But you can utilize Kaldi recipes in most cases
    echo "stage 1: Feature Generation"
    fbankdir=fbank
    # Generate the fbank features; by default 80-dimensional fbanks with pitch on each frame 对每一帧提取80维的fbank特征
    for x in test train; do
        steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 8 --write_utt2num_frames true \
            data/${
   x} exp/make_fbank/${
   x} ${
   fbankdir}
        utils/fix_data_dir.sh