Reconocimiento del habla con acento español basado en un modelo acústico

Abstract

The objective of the article was to generate an Automatic Speech Recognition (ASR) model based on the translation from human voice to text, being considered as one of the branches of artificial intelligence. Voice analysis allows identifying information about the acoustics, phonetics, syntax, semantics of words, among other elements where ambiguity in terms, pronunciation errors, similar syntax but different semantics can be identified, which represent characteristics of the language. The model focused on the acoustic analysis of words proposing the generation of a methodology for acoustic recognition from speech transcripts from audios containing human voice and the error rate per word was considered to identify the accuracy of the model. The audios were taken from the Integrated Security Service ECU911 that represent emergency calls registered by the entity. The model was trained with the CMUSphinx tool for the Spanish language without internet connection. The results showed that the word error rate varies in relation to the number of audios; that is, the greater the number of audios, the smaller number of erroneous words and the greater the accuracy of the model. The investigation concluded by emphasizing the duration of each audio as a variable that affects the accuracy of the model.