MARS: A Hybrid Deep CNN-based Multi-Accent Recognition System for English Language

Classifying the speech of non-native English speakers is challenging due to various features that distinguish accents. Accents vary by sex, age, formality, social status, geographical area, mother tongue, quality of the voice, phoneme, and prosody. This paper proposes a novel, well-structured database of non-native Indian English speaker accents, referred to as IndicAccentDB. IndicAccentDB contains speech samples from 6 different states to address the unbalanced dataset (gender-bias) and speaker mismatch problems observed in the past. The proposed work also discusses the requirements for creating the IndicAccentDB database and pre-processing tasks performed on the dataset. Furthermore, we experimented with accent classification models, namely 1D-CNN, Support Vector Machines, Random forest, Decision tree, ResNet18, ResNet50, and xResNet18, using MFCC and Mel-Spectrogram features to build the robust Multi-Accent Recognition System (MARS). At last, we evaluated the performance of proposed models on the novel database and compared the results using evaluation metrics like precision, accuracy, F1-score, and recall. Based on our findings, xResNet18 was able to identify the accent classes with significant accuracy.