Home

Awesome

狐呼标

中文文档

Intro

Use a neural network model to annotate the breathing (AP) in the textgrid file.

Note:

  1. Only label "AP" in the "SP" or "" label of the original tg file, and accuracy is based on the original annotation file.

  2. Only supports tg files with two layers of annotations: the words and the phones.

Cpp Version (Beta)

The CPP version has a UI interface, but cannot be accelerated using a graphics card.

How to use

  1. If using SOFA to generate textgrid annotations

    python infer.py ... --ap_detector NoneAPDetector

    An additional "--ap_detector NoneAPDetector" needs to be added to generate a tg file without AP annotations.

  2. Download model

    Model Link

    model_folder

    ├── config.yaml

    └── model_ckpt_steps_7000.ckpt

  3. Generate AP labels by running textgrid-add-ap

    python textgrid_add_ap.py --ckpt_path model_folder/xx.ckpt --wav_dir wav_dir --tg_dir tg_dir --tg_out_dir tg_out_dir

Option:
    --ckpt_path     str    Path to the checkpoint
    --wav_dir       str    Wav file folder (*.wav).
    --tg_dir        str    Textgrid files (*.TextGrid).
    --tg_out_dir    str    Output path of tg file after labeling AP.
    --ap_threshold  float  default: 0.4   Respiratory probability recognition threshold.  (Option)
    --ap_dur        float  default: 0.08  The shortest duration of breathing, discarded below this threshold, in seconds. (Option)
    --sp_dur        float  default: 0.1   SP fragments below this threshold will be adsorbed onto adjacent AP, in seconds.   (Option)

ReLabel

ReLabel the TG file with breathing.

  1. Clear AP in original label.

    python clean_ap.py --tg_dir raw_tg_dir --clean_tg_dir clean_tg_dir

 Option:
     --tg_dir        str    Textgrid files (*.TextGrid).
     --clean_tg_dir  str    Clean textgrid output dir (*.TextGrid).
     --phonemes      str    default: AP,SP,  The phonemes to be cleared are separated by English commas.  (Option)

  2. Generate AP labels by running textgrid-add-ap(to clean_tg_dir)