speech-aligner,是一个从“人声语音”及其“语言文本”,产生音素级别时间对齐标注的工具。
# 调用 bin,输入语音列表和文本、输出对齐结果
cd egs/cn_phn
speech-aligner --config=conf/align.conf data/wav.scp data/text data/out.ali
# 查看输出对齐结果,包含: 文件名,音素时间起点(秒) 音素时间终点(秒) 音素
cat data/text data/out.ali
BAC009S0002W0122 而对楼市成交抑制作用最大的限购
BAC009S0002W0122
0.000 0.535 sil
0.535 0.540 $0
0.540 0.745 er_2
0.745 0.850 d
0.850 0.895 ui_4
0.895 1.305 l
1.305 1.435 ou_2
...
4.955 5.055 x
5.055 5.525 ian_4
5.525 5.745 g
5.745 5.930 ou_4
5.930 5.975 sil
.
-
预先准备:
-
cmake >= 3.1
-
有如下blas接口数学库之一:
-
建议:mkl
- 安装 conda,并通过conda安装mkl:
conda install mkl
(mkl默认会随conda一起安装) - 编译时,确保
conda
可执行(which conda
有输出)
- 安装 conda,并通过conda安装mkl:
-
atlas
-
ubuntu安装:
sudo apt-get install libatlas3-base
-
linux发行版众多,数学库路径不一且变动,所以可以通过如下命令进行路径指定:
-
cmake -DBLAS_VENDORS=ATLAS -DBLAS_ATLAS_LIB_DIRS=[/path/to/atlas/lib ..
-
-
OSX系统(Darwin)自带Accelerate framework,可调过这项
-
…其他数学库,可查看
cmake/Modules/FindBLAS.cmake
,了解支持的数学库
-
-
-
cmake编译
git clone .../speech-aligner.git cd speech-aligner mkdir build && cd build cmake .. make -j
-
编译结果
- bin/speech-aligner: 二进制可执行文件,典型调用见egs/cn_phn/run.sh,包括三个参数:
- 配置:支持通过配置文件和命令行读取参数,建议使用如
--config=egs/cn_phn/conf/align.conf
- 输入:音频列表、对应的文本列表
- 输出:音素时间对齐标注
- 配置:支持通过配置文件和命令行读取参数,建议使用如
- bin/speech-aligner: 二进制可执行文件,典型调用见egs/cn_phn/run.sh,包括三个参数:
- 研究:
- 为TTS产生音素时间标注的训练数据
- egs/cn_phn
- 为TTS产生音素时间标注的训练数据
- 工程:
- 歌词对齐
- egs/cn_lyric [todo]
- 字幕对齐
- egs/cn_subtitle [todo]
- 歌词对齐
- for fun:
- 鬼畜
- egs/cn_gc [todo]
- 鬼畜
- 增加支持中文拼音(带调)输入,见egs/cn_phn/data/text
- 中文环境:标点和英文的处理
- 增加更多示例
speech-aligner, is a tool that generate phoneme-level alignment between human speech and its transcription
# call the bin,with speech and transcript as inputs
./bin/speech-aligner --config=egs/cn_phn/conf/align.conf egs/cn_phn/data/wav.scp egs/cn_phn/data/text egs/cn_phn/data/out.ali
# check the output alignment, include: filename, phoneme and its start/end time
cat egs/cn_phn/data/text egs/cn_phn/data/out.ali
BAC009S0002W0123
0.000 0.025 y
0.025 0.460 e_3
0.460 0.850 sil
0.850 0.985 ch
0.985 1.095 eng_2
...
2.655 2.735 zh
2.735 2.900 ong_1
2.900 2.960 d
2.960 3.665 ing_1
3.665 3.845 sil
.
-
requirements
-
cmake >= 3.1
-
one of blas math lib:
-
mkl (recommended)
- install conda, and use it to install mkl:
conda install mkl
(mkl is installed with conda by default) - when cmake,
conda
should be in your path
- install conda, and use it to install mkl:
-
atlas
-
ubuntu:
sudo apt-get install libatlas3-base
-
when cmake, it maynot find your atlas automatically, thus you need set the math lib path as below:
-
cmake -DBLAS_VENDORS=ATLAS -DBLAS_ATLAS_LIB_DIRS=[/path/to/atlas/lib ..
-
-
Accelerate framework (need do nothing for "macOS/Darwin")
-
...
-
-
-
cmake
git clone .../speech-aligner.git cd speech-aligner mkdir build && cd build cmake .. make -j
-
results
- bin/speech-aligner: a binary executable file, with arguments:
- configuration: through config file (recommendation, e.g.: --config=egs/cn_phn/conf/align.conf) or command line
- inputs: the wav list and the correspoing transcription list (e.g. egs/cn_phn/data )
- output: the result alignment
- bin/speech-aligner: a binary executable file, with arguments:
- for research:
- generate training data for TTS
- egs/cn_phn: generate chinese phoneme alignment
- generate training data for TTS
- for engineering:
- align lyric
- egs/cn_lyric [todo]
- align subtitle
- egs/cn_subtitle[todo]
- align lyric
- for fun:
- きちく
- egs/cn_gc [todo]
- きちく