Research Works

Romanized to Native Malayalam Script Transliteration

Developed a deep learning-based, general purpose, sentence level transliteration model to convert Romanized Malayalam text into native script, effectively handling various typing patterns. Demo. Paper

Exploring Pitfalls in Multilingual ASR Model Evaluations

Investigated text normalization issues in ASR models, uncovering biases in Indic language evaluations and proposing linguistically informed normalization techniques. This work accepted in EMNLP 2024 conference with an ACL grant for paper presentation.

Multistage Fine-tuning for Low-resource ASR

Developed an ASR model for Malasar, a low-resource Dravidian language, leveraging linguistic similarities with Tamil to enhance speech recognition performance despite data limitations. Demo. Paper.

Acoustic to Articulatory Inversion of Speech

Engineered a deep learning model combining BiLSTM and CNN to track tongue and lip articulatory features, contributing to advancements in speech production research.  Paper.

Evaluation of Tongue Articulation in Malayalam Vowels

Built an ANN-based system for analyzing vowel articulation in children’s speech, aiding in early diagnosis and intervention for speech disorders.  Paper.

Learning Assistive System for Autistic Children

Developed an ASR-based system to assess articulation disorders in children with Asperger’s Syndrome and designed an interactive speech training tool (LASAC) to aid learning through visual and auditory reinforcement.  Paper. 

Malayalam Vowel Classification in children

Investigated SVM-based vowel classification in children’s speech, optimizing performance through advanced kernel functions and hyperparameter tuning.  Paper..