Improving Performance of NeMo ASR system for Indian Accent English

While many speech recognition applications are quite commonplace now a days, applicability of the recognition techniques to accented languages is still a major challenge. Here, in this paper we have fine-tuned a state-of-the-art convolutional ASR system on regional accented Indian English to improve it’s performance for accented English. We used NVIDIA NeMo ¹ as our base framework and used QuartzNet 1D Convolutional model to explore their ability to understand language specific phonology. We also fine-tuned the convolutional network on regional accented Indian English. The tests were carried out on a dataset containing male and female English speech samples collected over 13 different regional accents. The experiments were bench-marked against the output generated by well known speech services and the effect of the convolution was found to be positive.