Performance evaluation of deep learning models for Indian-accented English automatic speech recognition
Аннотация
Automatic Speech Recognition (ASR) systems trained on native English accents often underperform for Indian-accented English due to distinct phonetic and prosodic variations. This study presents a comprehensive, architecture-controlled evaluation of acoustic modeling approaches for Indian-accented English ASR, leveraging a 126-hour annotated speech corpus (91 h from the NPTEL2020 Corpus and 35 h from a manually curated In-House Corpus) and a 45,000-word text corpus spanning six major Indian linguistic regions. We evaluate five acoustic-modeling systems in total: a conventional GMM-HMM baseline, refined progressively through four training stages (Mono-Phone, Tri-Phone, LDA+MLLT, and LDA+MLLT + SAT), and four discriminative neural architectures — DNN-HMM, TDNN, TDNN-BLSTM, and TDNN-LSTM-Attention — using Word Error Rate (WER) as the primary metric, and we additionally report a zero-shot, inference-only open-source Whisper (large-v3) baseline on the identical held-out test set to directly situate the Kaldi hybrid systems against a current open-source state-of-the-art model. All trained systems use identical 118.5-hour training / 7.5-hour test split of the combined corpus to isolate the effect of acoustic-model architecture. Results demonstrate that hybrid models (TDNN-BLSTM) achieve the lowest WER among the trained hybrid systems (8.22%), outperforming the best GMM-HMM configuration (24.23%) and standalone TDNN (10.31%). The TDNN-LSTM-Attention model (WER = 9.81%) further validates the utility of dynamic temporal weighting. We also analyze the insertion/deletion/substitution error composition itself as an interpretive tool, relating the dominance of substitution errors to documented segmental features of Indian English phonology. Our work highlights the necessity of accent-specific adaptations, revealing that speaker-adaptive transforms (SAT), dilated TDNNs, and bidirectional LSTMs are critical for modeling Indian-English phonetics. We further discuss, and explicitly justify, the study’s scope relative to recent self-supervised and transformer-based ASR systems (Whisper), positioning this work as a controlled hybrid-architecture benchmark for low-resource, on-premise deployment settings. The study establishes benchmarks for Indian-accented ASR within the Kaldi hybrid framework and provides a reproducible framework for developing inclusive speech technologies for non-native English varieties.