End-to-end ASR framework for Indian-English accent: using speech CNN-based segmentation

The superiority of Automatic Speech Recognition (ASR) has significantly enhanced over time, with a focus from short utterance circumstances to longer audio signal. In short utterances, speech endpoints are distinct, ensuring a good user experience. However, in long-form scenarios, these endpoints are less clear, leading to unnecessary resource consumption and deviating from ASR's primary goal of generating highly readable and well-formatted transcriptions. In this study, we introduced an ASR framework tailored for the Indian English accent. We employed Speech Segments Endpoint Detection (SSED), built using Mel-spectrogram features, short time energy signal, and a hybrid Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) model. Our experiments on a 29-h audio dataset containing Indian English accent speech achieved impressive results: the CNN–BiLSTM classification model for speech endpoint detection attained 98.67% accuracy in training and 93.62% accuracy in validation. The resulting ASR system achieved a Word Error Rate (WER) of 11.63%. Notably, the segmentation model reduced the dataset length by 16.4%, making it a valuable contribution to ASR technology. © 2023, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.