Front-end Based Robust Speech Recognition Methods: A Review
Vicky Zilvan, Ana Heryana, Asri Rizki Yuliani, Dikdik Krisnandi, Raden Sandra Yuwana, Hilman Ferdinandus Pardede · 2021
Improving the robustness of speech recognition against environmental noise for speech recognition has been a long research problem. Even with many developments in the field, it remains a notable challenge. A plethora of methods have been proposed to deal with these problems. Some of them deal with noise on the signal levels at the front-end of speech recognition while others adapt the model at the back-end to the noisy conditions. In the past, applying various signal processing techniques has been dominant for front-end-based models, while model adaptations are often applied at the back-end. Currently, data-driven based such as machine learning and deep learning have been increasingly more popular approaches. Both supervised and unsupervised methods have been employed for this task. In this study, our main objective is to review various challenges of solving the robustness issues of speech recognition against environmental noise. We focus on front-end-based methods since their improvements are more tractable for deep learning-based speech recognition. Furthermore, we also summarize and describe various techniques for robust speech recognition and the trends for future methods with the emphasis on data-driven-based methods. Finally, we discuss the advantages and disadvantages of these methods and report several important results in the field.