Authors:
Braulio Baldeon
;
Renzo Ravelli
and
Willy Ugarte
Affiliation:
Universidad Peruana de Ciencias Aplicadas (UPC), Lima, Peru
Keyword(s):
Phonological Processes, Language Development Level, Language Acquisition, Machine Learning.
Abstract:
In the context of the pandemic we are living in, most of the interactions between kindergarten-aged children has decreased, meaning that their language development might be slowed down. Our work presents a machine learning-based method for the classification of phonological processes and a corpus with a total of 3,324 audios, being 40% of them audios with a correct pronunciation, and the remaining 60% wrong. One of the main problems encountered when trying to perform children speech recognition in Spanish, is, to the best of our knowledge, the lack of a corpus. 329 audios were collected from 20-30 years old adults and a voice conversion technique was applied in order to generate the required audios for the corpus construction. A modified AlexNet was trained to classify if a word was correctly pronounced, if the audio was classified as mispronounced, it goes through a second AlexNet to classify which kind of Phonological Process is found in the word.