Speech Lab

Our work is centrally on ASR (Automatic Speech Recognition), spanning topics such as multi-lingual ASR with Indian languages applications, end-to-end (E2E) ASR architectures, attention mechanisms, unsupervised representation learning for low-resource ASR, analysis-by-synthesis frameworks for ASR, few-shot learning (FSL) for E2E ASR, associative memory formulations and multi-modal associative learning.

Our focus has been on examining various paradigms for ASR as indicated above, with an undercurrent guiding principle of utilizing and exploiting speech-knowledge within the current trends of highly data-hungry machine learning and deep learning approaches and frameworks. As an example, we refer to one of the recent works: ‘Few-shot learning’ (FSL) paradigms that utilize multi-lingual phone-sets and cross-lingual embeddings for effective decoding of target ultra low-resource languages.

We have developed ASR systems for several Indian languages such as Hindi, Marathi, Odia, Gujarati, Tamil, Telugu, Kannada, Malayalam, Bengali, Assamese, Urdu etc.