This repository consists of supprorting scripts for building datasets for ETTS in English
src
folder consists of scripts andconfigs
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).
- 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
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.
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 insource_data_directory
in config.yaml; - run emodb_dataset;
- place
cmuarctic.data
in the samesource_data_directory
.
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.