Abstract
Mother tongues and regional dialects have a substantial impact on pronunciation, leading to a range of complex and unique accents. This complexity increases in a diverse country such as India, which has code-mixed languages, which necessitates the development of an Automatic Speech Recognition(ASR) system capable of accommodating these variations effectively. In order to address this, we suggest a Hinglish voice recognition task that is cross-accented and uses the Hindi+English code-mixed conversational data to evaluate how well a model can adapt to different accents. The model-agnostic meta-learning (MAML) technique is extended by our accent-agnostic method to enable quick adaptation to unknown accents. We show through numerous experiments that our method works well. It outperforms joint multi-accent training in both mixed and cross-region settings, as measured by word error rate(WER), which is enhanced by 3-5%. Further, we investigate the effects of few-shot fine-tuning on mixed and cross-region samples, and ultimately revealing important insights into the pronunciation of Hindi-accented audio across different geographical areas of the country.