Noise reduction, dereverberation and binary neural networks for improved automatic speech recognition

Technologies using vocal commands are very useful in situations where hands cannot be used (e.g. wearing gloves or in factory settings to operate complex machines). The performance of automatic speech recognition systems decreases significantly in the presence of noise or reverberation (i.e. echoes on objects and walls). This projects aims at improving the performance of our partner’s automatic speech to intent recognition (ASIR) system by reducing environmental noise and reverberation. We will reduce noise by separating speech signals from noise signals, exploiting their independence. We will also develop signal processing methods to reduce the reverberation that distorts speech signals in a room. Our partner’s system must function without any internet connection and must also minimise its energy consumption. For this purpose, we will develop binary neural networks for ASIR. The methods developed will be integrated into a single system that will be designed to be optimized globally.

Faculty Supervisor:

Antoine Saucier;Jean-Pierre David

Student:

Salah Ben Slimene;Sameh Aissaoui

Partner:

Fluent.AI Inc

Discipline:

Engineering - computer / electrical

Sector:

University:

Polytechnique Montréal

Program: