Parturition Hindi Speech Dataset for Automatic Speech Recognition
Vansh Bansal, T Thishyan Raj, Nagarathna Ravi, Shubham Korde, Jaskaran Kalra, Sudha Murugesan, B Ramkrishnan, Aboli Gore, Vipul Arora · 2023
While automatic speech recognition (ASR) technologies have become mature, they are mostly being developed by industry for large scale commercial applications. There are many niche domains that can potentially benefit from ASR. These domains may need ASR for a specific limited vocabulary for a certain population with a distinct language and dialect. This paper details the complete procedure for developing an ASR solution for such an application. It presents Parturition Hindi Speech (PHS) dataset prepared for real-time ASR for a medical application in Bihar, India. The dataset is prepared for childbirth assistance with recordings done by nurses in situ. Finally, several ASR systems are developed for the PHS dataset and their performances are compared. The models are pre-trained on large datasets and are adapted to PHS dataset. In experiments, we find that end-to-end ASR models adapt more effectively as compared to GMM-HMM based models. Moreover, custom language models further boost the performance.