用 kaldi 和 CVTE开源模型 实现语音识别

下载模型

CVTE开源了kaldi的中文模型，
模型下载地址： http://kaldi-asr.org/models/0002_cvte_chain_model.tar.gz
解压放到kaldi/egs/下

使用

将egs/wsj/s5中的steps和utils拷贝到egs/cvte/s5目录下:
将egs/hkust/s5/local/score.sh拷贝到egs/cvte/s5/local目录下:

cp -r egs/wsj/s5/steps egs/cvte/s5/steps
cp -r egs/wsj/s5/utils egs/cvte/s5/utils
cp egs/hkust/s5/local/score.sh egs/cvte/s5/local

注释掉utils/lang/check_phones_compatible.sh中if语句中的exit 1：

 36 # check if the files exist or not37 if [ ! -f $table_first ]; then38   if [ ! -f $table_second ]; then39     echo "$0: Error! Both of the two phones-symbol tables are absent."40     echo "Please check your command"41     #exit 1;42   else43     # The phones-symbol-table1 is absent. The model directory maybe created by old script.44     # For back compatibility, this script exits silently with status 0.45     exit 0;46   fi

然后执行./run.sh就可以了

测试自己的数据集

准备文件

0. 音频文件

要求是16-bit位深，采样率16000Hz，单声道，wav格式的语言文件

1. wav.scp

wav.scp 格式

音频id 音频位置

如下：

AUDIO_20211129_170900_0000    ./audio/2021_11_29_17.09.00_0000.wav
AUDIO_20211129_170901_0000    ./audio/2021_11_29_17.09.01_0000.wav
AUDIO_20211129_170902_0000    ./audio/2021_11_29_17.09.02_0000.wav
AUDIO_20211129_170903_0000    ./audio/2021_11_29_17.09.03_0000.wav
AUDIO_20211129_170904_0000    ./audio/2021_11_29_17.09.04_0000.wav
AUDIO_20211129_170905_0000    ./audio/2021_11_29_17.09.05_0000.wav
AUDIO_20211129_170906_0000    ./audio/2021_11_29_17.09.06_0000.wav
AUDIO_20211129_170907_0000    ./audio/2021_11_29_17.09.07_0000.wav
AUDIO_20211129_170908_0000    ./audio/2021_11_29_17.09.08_0000.wav
AUDIO_20211129_170909_0000    ./audio/2021_11_29_17.09.09_0000.wav
AUDIO_20211129_170910_0000    ./audio/2021_11_29_17.09.10_0000.wav
AUDIO_20211129_170911_0000    ./audio/2021_11_29_17.09.11_0000.wav
AUDIO_20211129_170912_0000    ./audio/2021_11_29_17.09.12_0000.wav
AUDIO_20211129_170913_0000    ./audio/2021_11_29_17.09.13_0000.wav
AUDIO_20211129_170914_0000    ./audio/2021_11_29_17.09.14_0000.wav

2. utt2spk

音频ID
说话人ID
音频ID最好含有说话人ID

由于本例没有说话人，所以用音频ID代替说话人，即每条音频都是一个独立的说话人

utt2spk 格式

音频ID1 说话人1
音频ID2 说话人2

如下：

AUDIO_20211129_170900_0000    AUDIO_20211129_170900_0000
AUDIO_20211129_170901_0000    AUDIO_20211129_170901_0000
AUDIO_20211129_170902_0000    AUDIO_20211129_170902_0000
AUDIO_20211129_170903_0000    AUDIO_20211129_170903_0000
AUDIO_20211129_170904_0000    AUDIO_20211129_170904_0000
AUDIO_20211129_170905_0000    AUDIO_20211129_170905_0000
AUDIO_20211129_170906_0000    AUDIO_20211129_170906_0000
AUDIO_20211129_170907_0000    AUDIO_20211129_170907_0000
AUDIO_20211129_170908_0000    AUDIO_20211129_170908_0000
AUDIO_20211129_170909_0000    AUDIO_20211129_170909_0000
AUDIO_20211129_170910_0000    AUDIO_20211129_170910_0000
AUDIO_20211129_170911_0000    AUDIO_20211129_170911_0000
AUDIO_20211129_170912_0000    AUDIO_20211129_170912_0000
AUDIO_20211129_170913_0000    AUDIO_20211129_170913_0000
AUDIO_20211129_170914_0000    AUDIO_20211129_170914_0000

3. spk2utt

spk2utt 格式

说话人1 音频 音频 音频
说话人2 音频 音频 音频

有几个说话人就是几行，中间用空格隔开
如下：

AUDIO_20211129_170900_0000    AUDIO_20211129_170900_0000
AUDIO_20211129_170901_0000    AUDIO_20211129_170901_0000
AUDIO_20211129_170902_0000    AUDIO_20211129_170902_0000
AUDIO_20211129_170903_0000    AUDIO_20211129_170903_0000
AUDIO_20211129_170904_0000    AUDIO_20211129_170904_0000
AUDIO_20211129_170905_0000    AUDIO_20211129_170905_0000
AUDIO_20211129_170906_0000    AUDIO_20211129_170906_0000
AUDIO_20211129_170907_0000    AUDIO_20211129_170907_0000
AUDIO_20211129_170908_0000    AUDIO_20211129_170908_0000
AUDIO_20211129_170909_0000    AUDIO_20211129_170909_0000
AUDIO_20211129_170910_0000    AUDIO_20211129_170910_0000
AUDIO_20211129_170911_0000    AUDIO_20211129_170911_0000
AUDIO_20211129_170912_0000    AUDIO_20211129_170912_0000
AUDIO_20211129_170913_0000    AUDIO_20211129_170913_0000
AUDIO_20211129_170914_0000    AUDIO_20211129_170914_0000

测试：

替换data/fbank/test/ 下同名文件在 执行./run.sh就可以了

可见 准确率还是比较高的