Deep Learning System Based on the Separation of Audio Sources to Obtain the Transcription of a Conversation

Flores N.; Angeles D.; Tuesta S. · Journal of Advances in Information Technology (2022) · DOI: 10.12720/jait.13.4.393-397

Resumen

Podcasting has lately been in the spotlight for being the fastest-growing format, especially during the pandemic. This growth has highlighted the need for making podcasts accessible to diverse audiences, especially those having auditory disabilities. The current transcription methods have been unsatisfactory; therefore, we present an alternative method to transcribe audio files into text by segmenting audio sources. The applied methodology considers the construction of a public audio dataset having a duration of more than 15h. The training model was based on three scenarios in which the duration of the training data was varied to determine the best performance, which was 10.77 in terms of the scale-invariant signal-to-noise ratio. We have simplified podcasting accessibility by making available the source code of each component that we developed.

Instituciones: Universidad Nacional Mayor de San Marcos