Sign Language Translator

ISL is the primary form of communication for the
deaf and hard-of-hearing community in India, and until there
are automated translation tools to provide real-time
communication between ISL users and those who can hear,
the gap between these two groups will continue to widen. This
paper presents ISL-Net, a software-only end-to-end pipeline
that performs live recognition of ISL gestures and converts
them to spoken output without requiring any specialised
hardware. The system captures hand gestures through a
standard webcam or smartphone camera, extracts 21-point
3D skeletal landmarks per hand using Google Media Pipe,
and classifies dynamic sign sequences using a hybrid CNNLSTM
architecture trained on normalised landmark
coordinates. Recognised sign labels are assembled into
grammatically smoothed sentences and synthesised as Hindi
and English audio using the Gotts text-to-speech engine.
Experimental evaluation on a custom-recorded 10-class ISL
dataset demonstrates a top-1 classification accuracy of 91.4%,
with real-time inference operating at 18–28 frames per second
on standard laptop hardware. The proposed approach
eliminates dependence on pixel-level image features, making
it robust to variations in lighting, background and skin tone.
This work represents a step toward accessible, affordable
assistive communication technology adapted specifically for
Indian sign language conventions.