Overview
Artificial reverberation has been added to anechoic speech data for training more robust speech related machine learning models. We propose to use a novel geometric acoustic simulation method to model real-world speech more accurately and significantly reduce the performance gap between training on synthetic data and testing on real-world data. We have observed benefits in tasks including automated speech recognition (ASR), wake-up-word (WUW) spotting, and speech localization.
Software
pygsound: pygsound is a python package for impulse response generation based on state-of-the-art geometric sound propagation engine. The simulation is implemented with C++ and uses pybind11 for python APIs.
Datasets
6-channel synthetic impulse response dataset: Dataset used in our paper: Improving Reverberant Speech Training Using Diffuse Acoustic Simulation.
Low-frequency compensated synthetic impulse response dataset: Dataset used in our paper: Low-frequency Compensated Synthetic Impulse Responses for Improved Far-field Speech Recognition.