some-script-to-help-using-Montreal-Forced-Aligner

Some script for helping using Montreal Forced Aligner,

• mainly for transforming Hanzi character to pinyin,
• generating phone dict,
• train your own model
• and extrat pause time from .textgrid files.

In this project, I use Common Voice Chinese dataset as example. After extrating dataset.zip, please move clips directory to the path contains common-voice-dataset-hanzi2pinyin.py file. Then rename clips to data. Also make sure that zh-CN in that path.

Before traing with MFA(Montreal-Forced-Aligner)

1. 0_common-voice-dataset-hanzi2pinyin.py, will generate .lab files in data dir.
2. 1_trans_mp3_to_wav_and_move_mp3_to_another.py. Because Montreal-Forced-Aligner does not fit .mp3 for training, I transformed it to .wav and moved .mp3 to data_mp3.
3. Option. 1_copy_lab_to_data_mp3.py
4. 2_wav_subsample_to_16000.py, original .wav has many differnt speaker ratio. Using this scrip to change ratio.
5. 3_use_mfa_make_dict.sh, generate a dictionary of pinyin phone. Make sure you set the right path parameter in .sh.

Doing a training

4_use_mfa_train_align_model.sh, Training. Make sure you set the right path parameter in .sh.

After traing, to produce pause time

1. 5_1_process_textgrid_mid_as_sp.py, use sp and sil generated by MFA to extract pause time.
2. Or, 5_2_process_textgrid_only_sp.py, use the middle timepoint of the every character to compute a difference time of two neighbor character.