A transfer learning approach for continuous speech recognition system in Indian language sadri

A transfer learning approach for continuous speech recognition system in Indian language sadri

연구 분야: Artificial Intelligence

학회: International Journal of Information Technology

Sadri is the most widely used language of the Chotanagpur Plateau region of India. This is primarily a spoken language and developing an automatic speech recognition (ASR) system in Sadri is extremely important. When we searched the literature, we found no open ASR system or relevant resources in Sadri. So, we worked on developing an ASR system in Sadri. We created a Sadri speech corpus of around 20 h, employing 53 native speakers. Then, we implemented the baseline model using Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). As the size of the training data was not sufficient, we investigated a transfer learning technique to improve performance. We found that ASR data are openly available in other related languages such as Hindi and Bengali. We used Bengali and Hindi data of 100 h each in a transfer learning framework along with the Sadri data. In our experiments, we found that the transfer learning-based model outperforms the baseline model.