Aprendizaje profundo para selección de opciones numéricas por voz como herramientas para chatbot

Abstract

This document presents the design of a voice-operated chatbot-type assistant that works following a dialogue model between user and robot, which is trained with deep learning algorithms, using a database of spectrograms constructed from male and female voices, based on the short-time Fourier transform and Mel frequency cepstral coefficients as signal preprocessing techniques. For the recognition and classification of voice patterns, five convolutional network architectures are designed with the same parameters. The performance achieved in the training of the networks is compared, where all degrees of accuracy were greater than 92.8%. It is observed that the number of layers of the networks affects the number of learning parameters, their degree of accuracy and digital weight; in general, a greater number of layers increases both the training time and the classification time. Finally, for validation through a chatbot App, the selected network is applied to the completion of a survey that uses a Likert scale from 1 to 5, where users, in addition to saying the selected option, confirm it with a Yes or No, the App plays the audio of each question, shows its identification, listens and confirms the user's answers. The selected network design is concluded, allowing the development of chatbot applications based on audio interaction.