Procesamiento acústico de la voz aplicado al reconocimiento de fonemas similares en el habla inglesa comparados con sus homólogos en lengua hispánica : Evaluación de modelos basados en Transformer para la detección de errores de lectura a nivel de palabras.

Abstract

This study focuses on evaluating word-level reading errors in audio recordings of the Spanish language (NHK newscasts for Spanish) using the Transformer-based model: Whisper Large-V2. The evaluation results demonstrated the model's capacity to accurately predict most errors. Additionally, two experiments were conducted: the first involved an attempt to train a Transformer-based model using the Whisper-Small model through the transfer learning technique to predict phonetic-level pronunciation for the English language. The second experiment entailed assessing phoneme-level pronunciation errors using the "wav2vec2-xlsr-53-espeak-cv-ft" model for English, spoken by individuals whose native language is Spanish. Unfortunately, both experiments yielded unsatisfactory outcomes.