A Multi-Speaker Kannada Lip-Reading Video Dataset for Word and Phrase Recognition
Description
This dataset presents a multi-speaker Kannada lip-reading video corpus developed to support research in visual speech recognition for Kannada. The dataset consists of 27 Kannada utterance classes, including isolated words and short two-word phrases, recorded by eight different speakers. Each utterance class contains a minimum of 50 video samples collected across the speakers. Individual recordings have a duration of approximately 2–3 seconds or more and capture the visible facial and lip movements of the speaker while pronouncing the corresponding Kannada utterance. The dataset is organised speaker-wise. Each speaker directory contains 27 folders corresponding to the 27 Kannada utterance classes. This organisation allows the data to be used for both speaker-dependent and speaker-independent lip-reading experiments. The multi-speaker recordings provide variations in lip movement, facial characteristics, articulation, and speaking style. Speaker identities are represented using anonymous speaker identifiers. The dataset is intended for the development and evaluation of machine learning and deep learning methods for Kannada lip reading, visual speech recognition, and regional-language visual speech processing. Class-label and metadata files are provided to describe the utterances and associated video recordings.