Digital Life
Resources
Access to datasets: Libriparty and in-domain dataset
¿Te interesa?
Si eres profesor o estudiante universitario y tienes interés en participar en el programa TUTORÍA, registra tus datos para que podamos iniciar el programa.
Area of development
Proficiency in Python programming language and understanding of the fundamentals of machine learning are recommended for this project.
Challenge Intro
The popularity of voice-based interfaces has significantly increased due to their ability to
enable hands-free communication with a wide range of devices. In this context, deep learning technologies have emerged as the standard approach to
enhance naturalness and efficiency in device interactions.
Edge AI plays a crucial role in the development of such interfaces as it helps to minimize delays and protect user privacy. There are various approaches to improving communication between humans and devices. One of these approaches involves working at the signal level, specifically with speech audio. Typically, this signal is contaminated by the presence of background noise. The task of Voice Activity Detection aims to enhance users' interactions with devices by effectively distinguishing between unwanted noises and valid user utterances.
Edge AI plays a crucial role in the development of such interfaces as it helps to minimize delays and protect user privacy. There are various approaches to improving communication between humans and devices. One of these approaches involves working at the signal level, specifically with speech audio. Typically, this signal is contaminated by the presence of background noise. The task of Voice Activity Detection aims to enhance users' interactions with devices by effectively distinguishing between unwanted noises and valid user utterances.
Challenge Description
The research project objective is to evaluate existing Voice Activity Detection tools and integrate them into the human-device communication chain. The project can be divided into the following subtasks:
Researching the State-of-the-Art on Voice Activity Detection.
Select an existing tool with free commercial use availability based on the research.
Exploring and preprocessing an out-of-domain audio dataset: Libriparty.
Using that data to assess out-of-domain model performance.
Studying operating point selection and post-processing for real-time models.
Exploring and preprocessing an in-domain audio dataset: domestic scenario.
Using that data to assess in-domain model performance.
Selecting in-domain operating point and post-processing strategy.
Integrating the model into the human-device communication chain with virtual assistants.
Researching the State-of-the-Art on Voice Activity Detection.
Select an existing tool with free commercial use availability based on the research.
Exploring and preprocessing an out-of-domain audio dataset: Libriparty.
Using that data to assess out-of-domain model performance.
Studying operating point selection and post-processing for real-time models.
Exploring and preprocessing an in-domain audio dataset: domestic scenario.
Using that data to assess in-domain model performance.
Selecting in-domain operating point and post-processing strategy.
Integrating the model into the human-device communication chain with virtual assistants.
¿Quién te plantea este desafío?
Los Tutores Industriales de Telefónica, te acompañan en el desarrollo del TFG/TFM, aportando su visión real de la industria. Compartirán su conocimiento y experiencia, ofreciéndote feedback para que puedas desarrollar un proyecto con impacto innovador.
-1.jpg "Fernando López Telefónica")