dariadiatlova/emo-tts-data

ETTS data preprocessing 🌼

This repository consists of supprorting scripts for building datasets for ETTS in English

Repository structure:

• src folder consists of scripts and configs for processing different datasets;
• all data is processed to create uniform data structure in target_directory_path directory (path is set in relative config file).

Expected files structure:

- target_directory_path:
   - wavs:
      - {speaker_id}_{audio_id}_{emotion_id}.wav
      - ...
      - {speaker_id}_{audio_id}_{emotion_id}.txt
   - {dataset_name}_manifest_train.txt:
      - wav_path|speaker_id|emotion_id|transcription
   - {dataset_name}_manifest_evaluation.txt
   - {dataset_name}_manifest_test.txt

Datasets:

1. Parallel ESD

• Downloading link
• Arxiv

To process English-speakers data:

• fill source_data_directory in config.yaml with the root path to the dataset;
• run parallel_dataset;
• if script crashes with KeyError for some .wav file – open relative .txt file and add tab between audio-id and transcription.

2. EmovDB

• Downloading wavs, Download transcriptions
• Arxiv

We use only 2 emotions: Neutral, Angry taken from 3 speakers: Bea, Jenie, Sam. To process data:

• unarchive relevant tar.gz folders and point the directory in source_data_directory in config.yaml;
• run emodb_dataset;
• place cmuarctic.data in the same source_data_directory.

Profiler:

To get stats (emotion distribution per hour and total dataset size): fill profiler.yaml and run profiler.py.

Output:

    Emotion 'Neutral': 2.98 hours.
    Emotion 'Sad': 2.07 hours.
    Emotion 'Angry': 3.313 hours.
    Emotion 'Happy': 1.845 hours.
    Emotion 'Surprised': 1.869 hours.
    Total dataset size: 12.077 hours.