Deep learning based large vocabulary continuous speech recognition of an under-resourced language Bangladeshi Bangla

Abstract

Research in corpus-driven Automatic Speech Recognition (ASR) is advancing rapidly towards building a robust Large Vocabulary Continuous Speech Recognition (LVCSR) system. Under-resourced languages like Bangla require benchmarking large corpora for more research on LVCSR to tackle their limitations and avoid the biased results. In this paper, a publicly published large-scale Bangladeshi Bangla speech corpus is used to implement deep Convolutional Neural Network (CNN) based model and Recurrent Neural Network (RNN) based model with Connectionist Temporal Classification (CTC) loss function for Bangla LVCSR. In experimental evaluations, we find that CNN-based architecture yields superior results over the RNN-based approach. This study also emphasizes assessing the quality of an open-source large-scale Bangladeshi Bangla speech corpus and investigating the effect of the various high-order N-gram Language Models (LM) on a morphologically rich language Bangla. We achieve 36.12% word error rate (WER) using CNN-based acoustic model and 13.93% WER using beam search decoding with 5-gram LM. The findings demonstrate by far the state-of-the-art performance of any Bangla LVCSR system on a specific benchmarked large corpus.

BanSpeech: A Multi-Domain Bangla Speech Recognition Benchmark Toward Robust Performance in Challenging Conditions

“…Samin et al evaluated the quality of a large-scale publicly available LB-ASRTD corpus (229 hours) using deep learning-based approaches by conducting character-wise error analysis [20]. They also found a deep CNN-based acoustic model and a 5-gram Markov Language Model (LM) to be capable of achieving a lower word error rate (WER) on LB-ASRTD.…”

“…In this study, we also use a deep CNN-based model while utilizing a higher number of MFCCs during the input feature extraction and introducing layer normalization in each convolution layer. Based on an acoustic study on a regional accented speech and the character-wise error analysis on LB-ASRTD, the requirement of a new corpus with more speaker variability and character-wise well-balancedness was recommended [20], [21]. Therefore, Kibria et al developed the 241-hour-long publicly available Bangladeshi Bangla SUBAK.KO corpus with the aim of addressing the abovementioned issues of LB-ASRTD [7].…”

“…This is done for fully supervised CNN, weakly supervised Whisper, and selfsupervised wav2vec 2.0. While rescoring with the help of a language model achieves better performance [20], we would like to observe the WERs and CERs of the acoustic models without any interference from the language model. The goal of this study is to investigate the robustness of several ASR systems across multiple domains and evaluate SUBAK.KO on both read and spontaneous speech.…”

“…FIGURE 1.We use the same deep CNN architecture like[20] as our baseline in this study. The architecture consists of 20 convolutional layers, each followed by ReLU activation and a dropout layer.…”