Kaldi环境配置与Aishell训练

一、项目来源

代码来源：kaldi-asr/kaldi: kaldi-asr/kaldi is the official location of the Kaldi project. (github.com)

官网文档：Kaldi: The build process (how Kaldi is compiled) (kaldi-asr.org)

踩着我的同门@李思成-CSDN博客填上的坑kaldi环境配置与aishell实践_kaldi 编译-CSDN博客踏进新的坑里，Kaldi不是终端用户软件，也没有安装包，安装Kaldi包括编译其代码和准备必要的工具与运行环境。

二、环境配置

1、安装系统开发库的依赖

编译Kaldi之前，需要检查和安装Kaldi依赖的系统开发库，可以通过在Kaldi的tools目录下执行extras/check_dependencies.sh脚本来检查。

~/project/kaldi/tools$ extras/check_dependencies.sh
extras/check_dependencies.sh: subversion is not installed
extras/check_dependencies.sh: python2.7 is not installed
extras/check_dependencies.sh: Some prerequisites are missing; install them using the command:
  sudo apt-get install subversion python2.7

经过检查，我需要安装python2.7、python3和subversion

cd /home/chy524/kaldi/tools/
# 由于依赖库需要python2.7和python3，在tools目录下创建python文件夹，在文件夹中建立符号链接
mkdir -p python/

# 我还没有安装python2.7和python3，我在根目录下安装python2.7和python3
cd ~
mkdir -p ~/python

# 下载python2.7和python3.9
wget https://www.python.org/ftp/python/2.7.18/Python-2.7.18.tgz
wget https://www.python.org/ftp/python/3.9.0/Python-3.9.0.tgz

# 解压源码包
tar -xvzf Python-2.7.18.tgz
tar -xvzf Python-3.9.0.tgz

# 编译并安装，
cd Python-2.7.18
./configure --prefix=$HOME/python/python2.7
make
make install

cd ~/python_versions/Python-3.9.0
./configure --prefix=$HOME/python/python3
make
make install

# 安装完后，更新环境变量，打开.bashrc文件，在文件末尾添加以下内容
export PATH=$HOME/python/python2.7/bin:$PATH
export PATH=$HOME/python/python3/bin:$PATH

# 保存并关闭文件，然后执行
source ~/.bashrc

# 回到建立符号链接的python文件夹
cd /home/chy524/kaldi/tools/python/

# 建立符号链接
ln -s ~/python/python2.7/bin/python ~/python/python2.7/bin/python2
ln -s ~/python/python3/bin/python ~/python/python3/bin/python3

安装subversion，且我没有sudo权限

# 下载subversion
wget https://archive.apache.org/dist/subversion/subversion-1.14.1.tar.gz

# 解压源代码包
tar -xvzf subversion-1.14.1.tar.gz
cd subversion-1.14.1

# 安装 Subversion，我这里失败了，因为error: no suitable APR found，所以后面这些都执行不了
#./configure --prefix=$HOME/software/subversion
# make
# make install

# 将 Subversion 添加到环境变量中
# export PATH=$HOME/software/subversion/bin:$PATH

# 更新环境变量
# source ~/.bashrc

安装过程出现error: no suitable APR found报错，提示Subversion 在配置过程中没有找到所需的 Apache Portable Runtime (APR) 库。

configure: Apache Portable Runtime (APR) library configuration
checking for APR... no
configure: WARNING: APR not found
The Apache Portable Runtime (APR) library cannot be found.
Please install APR on this system and configure Subversion
with the appropriate --with-apr option.

You probably need to do something similar with the Apache
Portable Runtime Utility (APRUTIL) library and then configure
Subversion with both the --with-apr and --with-apr-util options.

configure: error: no suitable APR found

因此需要安装 APR 库以及 APR Utility（APRUTIL）库，才能成功编译和安装 Subversion，可以从 Apache APR 官方网站下载最新的源代码

# 下载APR与APRUTIL到~/software中
cd ~/software/
wget https://downloads.apache.org//apr/apr-1.7.0.tar.gz
wget https://downloads.apache.org//apr/apr-util-1.6.1.tar.gz

# 解压文件
tar -xvf apr-1.7.0.tar.gz
tar -xvf apr-util-1.6.1.tar.gz

# 编译并安装APR与APRUTIL
cd apr-1.7.0
./configure --prefix=$HOME/software/apr
make
make install

cd ../apr-util-1.6.1
./configure --prefix=$HOME/software/apr-util --with-apr=$HOME/software/apr

# 在apr-util中进行make时，出现问题：缺少了expat.h头文件，所以这后面执行不了
# make
# make install

在apr-util中进行make时，出现问题：缺少了expat.h头文件

-I/home/chy524/software/apr-util-1.6.3/include/private  -I/home/chy524/software/apr/include/apr-1    -o xml/apr_xml.lo -c xml/apr_xml.c && touch xml/apr_xml.lo 
xml/apr_xml.c:35:10: fatal error: expat.h: No such file or directory
   35 | #include <expat.h>
      |          ^~~~~~~~~
compilation terminated.
make[1]: *** [/home/chy524/software/apr-util-1.6.3/build/rules.mk:207: xml/apr_xml.lo] Error 1
make[1]: Leaving directory '/s6home/chy524/software/apr-util-1.6.3'
make: *** [/home/chy524/software/apr-util-1.6.3/build/rules.mk:119: all-recursive] Error 1

expat.h是Expat XML 解析库的头文件，而 APR-UTIL 在某些功能（比如 XML 处理）中依赖于它，所以需要安装Expat库

# 安装Expat到software中
cd ~/software/
wget https://github.com/libexpat/libexpat/releases/download/R_2_4_9/expat-2.4.9.tar.bz2

# 解压和编译Expat
tar -xvf expat-2.4.9.tar.bz2
cd expat-2.4.9

# 编译并安装到用户目录
./configure --prefix=$HOME/software/expat/
make
make install

# 配置APR-UTIL使用Expat
cd apr-util-1.6.3
./configure --prefix=$HOME/software/apr-util --with-apr=$HOME/software/apr --with-expat=$HOME/software/expat
make
make install

随后又回到subversion进行安装，出现了缺失serf库和sqlite3库，就不贴报错了，跟前面差不多，我继续解决报错了。安装serf库，从官网下载Download Apache Serf Sources，serf 是一个库，而不是可执行程序，所以没有 bin/ 目录。

# 安装serf库
cd ~/software/
wget https://archive.apache.org/dist/serf/serf-1.3.10.tar.bz2
tar -xvf serf-1.3.10.tar.bz2
cd serf-1.3.10

# 编译serf库，源代码中没有configure文件，不能像上面那样很简单就编译，只能通过提供的的readme文档进行编译，serf 使用的是 SCons 构建系统，而不是传统的 configure 脚本，可以从文件名 SConstruct 看出来，SConstruct 是 SCons 用来定义构建规则的文件。

# 下载 scons-local
mkdir -p ~/software/scons-local
cd ~/software/scons-local
wget http://prdownloads.sourceforge.net/scons/scons-local-2.3.0.tar.gz
tar -xvzf scons-local-2.3.0.tar.gz

# 检查安装是否成功
scons --version

# 构建Apache Serf
# 我的OPENSSL通过openssl version -a进行路径查找，在/home/chy524/miniconda3/ssl中（如果没有就下载OPENSSL），修改OPENSSL的路径
cd ~/software/serf-1.3.10
scons APR=$HOME/software/apr \
      APU=$HOME/software/apr-util \
      OPENSSL=$HOME/miniconda3 \
      PREFIX=$HOME/software/serf

# 安装Serf
scons PREFIX=$HOME/software/serf install
echo 'export PATH=$HOME/software/serf/bin:$PATH' >> ~/.bashrc
source ~/.bashrc

安装sqlite3库

cd ~/software/
# 下载 SQLite 源码
wget https://www.sqlite.org/2015/sqlite-amalgamation-3081101.zip
# 解压
unzip sqlite-amalgamation-3081101.zip
# 重命名文件夹为 sqlite-amalgamation
mv sqlite-amalgamation-3081101 /home/chy524/software/subversion-1.14.1/sqlite-amalgamation
# sqlite3就下载好了

再次回去编译subversion，发现找不到 LZ4 库，根据提示信息，直接修改参数就好,加一行--with-lz4=internal

cd ~/software/subversion-1.14.1
./configure --prefix=$HOME/software/subversion --with-apr=$HOME/software/apr --with-apr-util=$HOME/software/apr-util --with-serf=$HOME/software/serf --with-sqlite=$HOME/software/sqlite-amalgamation 
make
# make install

......
......
checking for inflate in -lz... yes
configure: lz4 configuration without pkg-config
checking for LZ4_compress_default in -llz4... no
configure: error: Subversion requires LZ4 >= r129, or use --with-lz4=internal

后面还出现一个库也需要改为内部库，直接加上就好，然后没有错误之后，最终执行安装好所有的前置配置，最终安装命令为

cd ~/software/subversion-1.14.1
./configure --prefix=$HOME/software/subversion --with-apr=$HOME/software/apr --with-apr-util=$HOME/software/apr-util --with-serf=$HOME/software/serf --with-sqlite=$HOME/software/sqlite-amalgamation --with-lz4=internal --with-(这里还有一个什么忘记了)=internal
make
make install

到此就完成了安装依赖的系统开发库

2、安装依赖的第三方工具库

无法访问GitHub，需要下载第三方工具库压缩包到tools目录下，并在tools/Makefile中调整库的版本号，然后执行make命令进行编译。

压缩包链接：https://pan.baidu.com/s/1la0nmku2zBxa1lRKiUTWcw?pwd=flho

修改tools/Makefile中的版本号：

# SHELL += -x

CXX ?= g++
CC ?= gcc        # used for sph2pipe
# CXX = clang++  # Uncomment these lines...
# CC = clang     # ...to build with Clang.

WGET ?= wget

OPENFST_VERSION ?= 1.7.2
CUB_VERSION ?= 1.8.0
# No '?=', since there exists only one version of sph2pipe.
SPH2PIPE_VERSION = 2.5
# SCTK official repo does not have version tags. Here's the mapping:
# 2.4.9 = 659bc36; 2.4.10 = d914e1b; 2.4.11 = 20159b5.
SCTK_GITHASH = 2.4.12

执行make进行编译

cd ~/project/kaldi/tools/
make

3. 安装语音模型工具IRSTLM / SRILM / Kaldi_lm

安装irstlm.sh

cd ~/project/kaldi/tools/
# 将install_irstlm.sh中https://github.com/irstlm-team/irstlm.git改为https://gitclone.con/github.com/irstlm-team/irstlm.git
extras/install_irstlm.sh
# 安装完成,加载环境变量
source ../tools/env.sh

安装kaldi_lm.sh

cd ~/project/kaldi/tools/
# 将install_kaldi_lm.sh中https://github.com/danpovey/kaldi_lm.git改为https://gitclone.com/github.com/danpovey/kaldi_lm.git
extras/install_irstlm.sh
# 安装完成,加载环境变量
source ../tools/env.sh

安装srilm.sh，使用手动安装

# 安装srilm之前先安装依赖lbfgs库
cd ~/project/kaldi/tools/
extras/install_liblbfgs.sh

# 在tools/目录下载压缩包
# https://pan.baidu.com/s/1la0nmku2zBxa1lRKiUTWcw?pwd=flho

# 解压并改名
tar -xvzf srilm-1.7.3.tar.gz
mv srilm-1.7.3/ srilm/

#将下方shell内容复制粘贴到install_srilm.sh中，随后运行
extras/install_srilm.sh

# 安装完成,加载环境变量
source ../tools/env.sh

修改install_srilm.sh中内容

#!/usr/bin/env bash

current_path=`pwd`
current_dir=`basename "$current_path"`

if [ "tools" != "$current_dir" ]; then
    echo "You should run this script in tools/ directory!!"
    exit 1
fi

if [ ! -d liblbfgs-1.10 ]; then
    echo Installing libLBFGS library to support MaxEnt LMs
    bash extras/install_liblbfgs.sh || exit 1
fi

! command -v gawk > /dev/null && \
   echo "GNU awk is not installed so SRILM will probably not work correctly: refusing to install" && exit 1;

# Skip SRILM download and extraction
# if [ ! -f srilm.tgz ] && [ ! -f srilm.tar.gz ] && [ ! -d srilm ]; then
#   if [ $# -ne 4 ]; then
#       echo "SRILM download requires some information about you"
#       echo
#       echo "Usage: $0 <name> <organization> <email> <address>"
#       exit 1
#   fi

#   srilm_url="http://www.speech.sri.com/projects/srilm/srilm_download2.php"
#   post_data="file=1.7.3&name=$1&org=$2&email=$3&address=$4&license=on"

#   if ! wget --post-data "$post_data" -O ./srilm.tar.gz "$srilm_url"; then
#       echo 'There was a problem downloading the file.'
#       echo 'Check your internet connection and try again.'
#       exit 1
#   fi

#   if [ ! -s srilm.tar.gz ]; then
#       echo 'The file is empty. There was a problem downloading the file.'
#       exit 1
#   fi
# fi

mkdir -p srilm
cd srilm

# Skip extraction of SRILM package
# if [ -f ../srilm.tgz ]; then
#     tar -xvzf ../srilm.tgz || exit 1 # Old SRILM format
# elif [ -f ../srilm.tar.gz ]; then
#     tar -xvzf ../srilm.tar.gz || exit 1 # Changed format type from tgz to tar.gz
# fi

# Ensure SRILM directory exists
# if [ ! -d srilm ]; then
#     echo 'The SRILM directory does not exist. There was a problem with the extraction.'
#     exit 1
# fi

# if [ ! -f RELEASE ]; then
#     echo 'The file RELEASE does not exist. There was a problem extracting.'
#     exit 1
# fi

major=`gawk -F. '{ print $1 }' RELEASE`
minor=`gawk -F. '{ print $2 }' RELEASE`
micro=`gawk -F. '{ print $3 }' RELEASE`

if [ $major -le 1 ] && [ $minor -le 7 ] && [ $micro -le 1 ]; then
  echo "Detected version 1.7.1 or earlier. Applying patch."
  patch -p0 < ../extras/srilm.patch
fi

# set the SRILM variable in the top-level Makefile to this directory.
cp Makefile tmpf

cat tmpf | gawk -v pwd=`pwd` '/SRILM =/{printf("SRILM = %s\n", pwd); next;} {print;}' \
  > Makefile || exit 1
rm tmpf

mtype=`sbin/machine-type`

echo HAVE_LIBLBFGS=1 >> common/Makefile.machine.$mtype
grep ADDITIONAL_INCLUDES common/Makefile.machine.$mtype | \
    sed 's|$| -I$(SRILM)/../liblbfgs-1.10/include|' \
    >> common/Makefile.machine.$mtype

grep ADDITIONAL_LDFLAGS common/Makefile.machine.$mtype | \
    sed 's|$| -L$(SRILM)/../liblbfgs-1.10/lib/ -Wl,-rpath -Wl,$(SRILM)/../liblbfgs-1.10/lib/|' \
    >> common/Makefile.machine.$mtype

make || exit

cd ..
(
  [ ! -z "${SRILM}" ] && \
    echo >&2 "SRILM variable is aleady defined. Undefining..." && \
    unset SRILM

  [ -f ./env.sh ] && . ./env.sh

  [ ! -z "${SRILM}" ] && \
    echo >&2 "SRILM config is already in env.sh" && exit

  wd=`pwd`
  wd=`readlink -f $wd || pwd`

  echo "export SRILM=$wd/srilm"
  dirs="\${PATH}"
  for directory in $(cd srilm && find bin -type d ) ; do
    dirs="$dirs:\${SRILM}/$directory"
  done
  echo "export PATH=$dirs"
) >> env.sh

echo >&2 "Installation of SRILM finished successfully"
echo >&2 "Please source the tools/env.sh in your path.sh to enable it"

4. 安装线性代数库OpenBLAS，不可访问github的情况

cd ~/project/kaldi/tools/

# 下载OPENBLAS_VERSION进行手动编译
git clone https://github.com/xianyi/OpenBLAS.git
cd OpenBLAS

# 适配 CPU 核心数并加速编译
make -j 8
make PREFIX=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install install

# 添加到环境变量
echo 'export LD_LIBRARY_PATH=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install/lib:$LD_LIBRARY_PATH' >> ~/.bashrc
echo 'export LIBRARY_PATH=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install/lib:$LIBRARY_PATH' >> ~/.bashrc
echo 'export C_INCLUDE_PATH=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install/include:$C_INCLUDE_PATH' >> ~/.bashrc
source ~/.bashrc

5.安装CUDA11.6，代码实践——准备阶段-CSDN博客有安装CUDA12.1的步骤，感觉还是11.6用的比较多

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Tue_Mar__8_18:18:20_PST_2022
Cuda compilation tools, release 11.6, V11.6.124
Build cuda_11.6.r11.6/compiler.31057947_0

创建运行所需的虚拟环境

# 创建conda虚拟环境
conda create -n kaldi python=3.8
conda activate kaldi
# 安装Pytorch
conda install cudatoolkit=11.6 -c pytorch -c conda-forge

6.编译Kaldi代码

cd ~/project/kaldi/src
# --cudatk-dir和--openblas-root需要为自己的地址，--cuda-arch=-arch需要自己的GPU版本
./configure --shared --use-cuda=yes --cudatk-dir=/home/chy524/cuda/cuda-11.6/ --cuda-arch=-arch=sm_86 --mathlib=OPENBLAS --openblas-root=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install/

# 处理 Kaldi 项目中所有的依赖关系，-j 8 参数可以让 make 使用 8 个并行线程加快这个过程。
make depend -j 8 

# 开始编译 Kaldi
make -j 8

三、跑Aishell数据集

1、数据准备部分，这些在终端都可以直接运行

# s5是语音识别模块，s5中的run.sh包含了这一模块的所有流程和代码，这是数据准备部分

 data=/home/chy524/data/aishell # 改成自己存放数据的地址
 data_url=www.openslr.org/resources/33

 . ./cmd.sh

# 我之前下载好了数据，就没运行这个
local/download_and_untar.sh $data $data_url data_aishell || exit 1;
local/download_and_untar.sh $data $data_url resource_aishell || exit 1;

# 准备词典
local/aishell_prepare_dict.sh $data/resource_aishell || exit 1;

# 准备数据,
local/aishell_data_prep.sh $data/data_aishell/wav $data/data_aishell/transcript || exit 1;

# 创建一个包含语言模型和词典的语言目录
# --position-dependent-phones false：指定是否使用与位置相关的音素。
# data/local/dict：包含字典的目录，字典包含词到音素的映射。
# "<SPOKEN_NOISE>"：用于表示未在字典中定义的噪声或其他非词汇性项。
# data/local/lang：中间输出目录。
# data/lang：最终的语言目录输出
utils/prepare_lang.sh --position-dependent-phones false data/local/dict \
    "<SPOKEN_NOISE>" data/local/lang data/lang || exit 1;

2、语言模型（LM）训练，这个在终端可以直接运行，很快

local/aishell_train_lms.sh

3、语言模型的格式化，这个在终端可以直接运行，很快

使用 utils/format_lm.sh 将语言模型（LM）转换为所需的 FST 格式，并生成 data/lang_test。

utils/format_lm.sh data/lang data/local/lm/3gram-mincount/lm_unpruned.gz \
    data/local/dict/lexicon.txt data/lang_test

4、生成 MFCC 和音高特征，这个开始需要在run.sh中注释掉其他部分代码，训练会生成文件及日志，建议进行提交训练，因为需要时间蛮久的

mfccdir=mfcc
for x in train dev test; do
  steps/make_mfcc_pitch.sh --cmd "$train_cmd" --nj 10 data/$x exp/make_mfcc/$x $mfccdir || exit 1;
  steps/compute_cmvn_stats.sh data/$x exp/make_mfcc/$x $mfccdir || exit 1;
  utils/fix_data_dir.sh data/$x || exit 1;
done

exp/make_mfcc/train：训练集日志

exp/make_mfcc/dev：验证集日志

exp/make_mfcc/test：测试集日志

后面我都没有跑了，CPU跑任务都太久了，不过有同门的博客可以学习kaldi环境配置与aishell实践_kaldi 编译-CSDN博客https://blog.csdn.net/weixin_46560570/article/details/141109113?spm=1001.2014.3001.5502

5、单音素模型

# 注释run.sh其他内容，运行./run.sh
# Train a monophone model on delta features.
steps/train_mono.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/mono || exit 1;

# Decode with the monophone model.
utils/mkgraph.sh data/lang_test exp/mono exp/mono/graph || exit 1;
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/mono/graph data/dev exp/mono/decode_dev
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/mono/graph data/test exp/mono/decode_test

# Get alignments from monophone system.
steps/align_si.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/mono exp/mono_ali || exit 1;

6、三音素模型（Tri1）

# 注释run.sh其他内容，运行./run.sh
# Train the first triphone pass model tri1 on delta + delta-delta features.
steps/train_deltas.sh --cmd "$train_cmd" \
 2500 20000 data/train data/lang exp/mono_ali exp/tri1 || exit 1;

# decode tri1
utils/mkgraph.sh data/lang_test exp/tri1 exp/tri1/graph || exit 1;
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/tri1/graph data/dev exp/tri1/decode_dev
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/tri1/graph data/test exp/tri1/decode_test

# align tri1
steps/align_si.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/tri1 exp/tri1_ali || exit 1;

7、更高级的三音素模型（Tri2）

# train tri2 [delta+delta-deltas]
steps/train_deltas.sh --cmd "$train_cmd" \
 2500 20000 data/train data/lang exp/tri1_ali exp/tri2 || exit 1;

# decode tri2
utils/mkgraph.sh data/lang_test exp/tri2 exp/tri2/graph
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/tri2/graph data/dev exp/tri2/decode_dev
steps/decode.sh --cmd "$decode_cmd" --config conf/decode.config --nj 10 \
  exp/tri2/graph data/test exp/tri2/decode_test

# Align training data with the tri2 model.
steps/align_si.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/tri2 exp/tri2_ali || exit 1;

8、Tri3a 模型

# Train the second triphone pass model tri3a on LDA+MLLT features.
steps/train_lda_mllt.sh --cmd "$train_cmd" \
 2500 20000 data/train data/lang exp/tri2_ali exp/tri3a || exit 1;
 
# Run a test decode with the tri3a model.
utils/mkgraph.sh data/lang_test exp/tri3a exp/tri3a/graph || exit 1;
steps/decode.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
  exp/tri3a/graph data/dev exp/tri3a/decode_dev
steps/decode.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
  exp/tri3a/graph data/test exp/tri3a/decode_test

# align tri3a with fMLLR
steps/align_fmllr.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/tri3a exp/tri3a_ali || exit 1;

9 、Tri4a模型

# Train the third triphone pass model tri4a on LDA+MLLT+SAT features.
# From now on, we start building a more serious system with Speaker
# Adaptive Training (SAT).
steps/train_sat.sh --cmd "$train_cmd" \
  2500 20000 data/train data/lang exp/tri3a_ali exp/tri4a || exit 1;

# decode tri4a
utils/mkgraph.sh data/lang_test exp/tri4a exp/tri4a/graph
steps/decode_fmllr.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
  exp/tri4a/graph data/dev exp/tri4a/decode_dev
steps/decode_fmllr.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
  exp/tri4a/graph data/test exp/tri4a/decode_test
  
# align tri4a with fMLLR
steps/align_fmllr.sh  --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/tri4a exp/tri4a_ali

10、Tri5a模型

# Train tri5a, which is LDA+MLLT+SAT
# Building a larger SAT system. You can see the num-leaves is 3500 and tot-gauss is 100000

steps/train_sat.sh --cmd "$train_cmd" \
  3500 100000 data/train data/lang exp/tri4a_ali exp/tri5a || exit 1;
  
# decode tri5a
utils/mkgraph.sh data/lang_test exp/tri5a exp/tri5a/graph || exit 1;
steps/decode_fmllr.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
   exp/tri5a/graph data/dev exp/tri5a/decode_dev || exit 1;
steps/decode_fmllr.sh --cmd "$decode_cmd" --nj 10 --config conf/decode.config \
   exp/tri5a/graph data/test exp/tri5a/decode_test || exit 1;
   
# align tri5a with fMLLR
steps/align_fmllr.sh --cmd "$train_cmd" --nj 10 \
  data/train data/lang exp/tri5a exp/tri5a_ali || exit 1;

11、nnet3模型

修改配置文件

# 若提示
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.
# 解决方法
cd kaldi/src/
修改configure文件
# 将你的/usr/local/cuda /home/chy524/cuda/cuda-11.6地址添加进去
function configure_cuda {
  # Check for CUDA toolkit in the system
  if [ ! -d "$CUDATKDIR" ]; then
    for base in /usr/local/share/cuda /usr/local/cuda /home/chy524/cuda/cuda-11.6 /usr/; do
      if [ -f $base/bin/nvcc ]; then
        CUDATKDIR=$base
      fi
    done
  fi
# 再次编译kaldi
./configure --shared --use-cuda=yes --cudatk-dir=/home/chy524/cuda/cuda-11.6/ --cuda-arch=-arch=sm_86 --mathlib=OPENBLAS --openblas-root=/home/chy524/project/kaldi/tools/OpenBLAS-0.3.13/install/

make -j clean depend
make -j 16

正式开始前需要将s5/local/nnet3/run_tdnn.sh中 --use-gpu true 修改为 --use-gpu wait

# nnet3
local/nnet3/run_tdnn.sh

12、chain模型

# chain
local/chain/run_tdnn.sh

13、得到结果

# getting results (see RESULTS file)
for x in exp/*/decode_test; do [ -d $x ] && grep WER $x/cer_* | utils/best_wer.sh; done 2>/dev/null
for x in exp/*/*/decode_test; do [ -d $x ] && grep WER $x/cer_* | utils/best_wer.sh; done 2>/dev/null