The most basic need for training deep learning models is to have suitable data, the more comprehensive and voluminous these data are, the more complete the learning of the models will be.
Based on this, this progect introduces the data collected and used in the training and evaluation of speech activity recognition models.
In this project, ERI_VADS-V01 dataset designed for training and evaluation of speech activity recognition models focusing on Persian language is introduced.
The ERI_VADS-V01 clean dataset consists of the cleaned Common Voice dataset and a portion of the TIMIT dataset assigned to 20-ms long, speech- and non-speech-labeled frames.
The TIMIT dataset has, in fact, been added to cover the English language. ERI_VADS-V01 noise dataset includes QUT, MUSAN and Audioset datasets.
To prepare the ERI_VADS-V01 dataset, the clean data is randomly combined with one or two noise data with a specified SNR.
In this dataset, to use reverberation, different reverberations are added to clean data and noisy data, and then they are summed together with a certain SNR. In the ERI_VADS-V01 dataset, 20% of the evaluation and test data includes reverberation, and in training, due to the fact that reverberation with noise can disrupt training, reverberation was not used at first, and then to complete the training, Resonance was used for 20% of the data. It should be noted that the training data is created online in order to increase the diversity of the data, but the evaluation and test data are prepared separately from the clean, noise and reverberation data of the training, once prepared and then stored so that the results of the models remain constant on the evaluation and test data.