speech aligner

speech-aligner

Chinese readme:

speech-aligner is a tool that generates phoneme-level time alignment annotations from "vocal voice" and its "language text".

Example

 # 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
.

Compilation

• Prepare:

  • cmake >= 3.1

  • There is one of the following Blas interface math libraries:

    • Suggestions: mkl

      • Install conda and install mkl through conda: conda install mkl (mkl will be installed with conda by default)
      • When compiling, make sure that conda is executable ( which conda has output)

    • atlas

      • ubuntu installation: sudo apt-get install libatlas3-base

      • There are many Linux distributions, and the paths of the math library are different and change, so you can specify the path through the following command:

      • cmake -DBLAS_VENDORS=ATLAS -DBLAS_ATLAS_LIB_DIRS=[/path/to/atlas/lib ..

    • OSX system (Darwin) comes with Accelerate framework, which can be adjusted

    • …Other math libraries, you can view cmake/Modules/FindBLAS.cmake to learn about the supported math libraries

• cmake compilation

  git clone .../speech-aligner.git
  cd speech-aligner
  mkdir build && cd build
  cmake ..
  make -j

• Compilation results

  • bin/speech-aligner: A binary executable file, typical calls are shown in egs/cn_phn/run.sh, including three parameters:
    • Configuration: Supports reading parameters through configuration files and command lines. It is recommended to use such as --config=egs/cn_phn/conf/align.conf
    • Input: Audio list, corresponding text list
    • Output: Phoneme time alignment annotation

Application scenarios and examples

• Research:
  • Generate phoneme time-noted training data for TTS
    • egs/cn_phn
• project:
  • Lyrics Alignment
    • egs/cn_lyric [todo]
  • Subtitle alignment
    • egs/cn_subtitle [todo]
• for fun:
  • Ghost beast
    • egs/cn_gc [todo]

renew

• Added support for Chinese pinyin (with tune) input, see egs/cn_phn/data/text

Todo

• Chinese environment: punctuation and English processing
• Add more examples

about

• The project is based on the famous open source project of voice kaldi , which follows the original project.
• Example egs/cn_phn, the phoneme list used is from another Chinese dictionary open source project DaCiDian .

---

English readme:

speech-aligner, is a tool that generate phoneme-level alignment between human speech and its transcription

Usage example

 # 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
.

Compile

• 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

    • atlas

      • ubuntu: sudo apt-get install libatlas3-base

      • When cmake, it may not 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, eg: --config=egs/cn_phn/conf/align.conf) or command line
    • inputs: the wav list and the correlation transcription list (eg egs/cn_phn/data )
    • output: the result alignment

Applications

• for research:
  • generate training data for TTS
    • egs/cn_phn: generate chinese phoneme alignment
• for engineering:
  • align lyric
    • egs/cn_lyric [todo]
  • align subtitle
    • egs/cn_subtitle[todo]
• for fun:
  • きちく
    • egs/cn_gc [todo]

About

• This project is based on a great speech open-source project kaldi .
• The phonemes used in the environment: egs/cn_phn, come from a chinese dictionary open-source project DaCiDian .