BHOSAI/PseudoLabelled_Azerbaijani_Voices

Azerbaijani dataset for training Automatic Speech Recognition

We are delighted to present voice set which contains more than 200k voice samples collected from youtube videos which are carefully selected in Azerbaijani language. Dataset is designed for automatic speech recognition tasks without labels. Dataset contains pseudo labels generated from Whisper-large-v3 model by setting the language to 'az'.

Steps below has been done in order to create the dataset:

1. Full Youtube video has been derived from the source.
2. The first audio of 20 seconds is derived from the given audio
3. 20 seconds of audio is given to whisper model and end timestamp of 1st sentence is used to get start of the second audio.
4. Loop continues until the audio is fully divided into the segments

At the end we get the audio per sentence. After having this set, we use whisper model again in order to get necessary pseudo labels.