Single-Frame Recognition of the Kazakh Sign Language Fingerspelling Alphabet: The Open QazSL Dataset and a Compact Hybrid Architecture

Kazakh Sign Language (KSL) has received little attention in automatic sign language recognition. Its fingerspelling alphabet contains 44 signs (42 letters and two function signs), and 12 of the letters (28.6%) involve hand movement in addition to handshape. This paper presents QazSL, a dataset covering all 44 classes, with 110,000 labelled video recordings (about 2500 per class, red–green–blue (RGB), 640 × 480) made by five native signers. An optimal-frame selection module extracted one 50 × 50 greyscale image per recording, which defines a deliberately simplified single-frame classification task: letters with movement are recognised from the handshape held during the sign, not from their movement. Eight architectures were compared under identical training conditions, from a baseline convolutional neural network (CNN) to attention-based, recurrent and multi-scale Inception models with residual connections. The proposed Inception Hybrid model (2.85 million parameters) achieves 98.49% accuracy on a validation split that shares signers with the training data and 97.41% on an independent test set of five new signers, in single training runs. All 44 classes are recognised with at least 95% accuracy; the lowest values are obtained for letters with movement, which are most often confused with static letters of similar handshape, showing the limits of the single-frame approach. On a consumer laptop, inference takes about 1.5 ms per frame on the graphics processing unit (GPU) and 41–43 ms on one thread of the central processing unit (CPU). Only an ordinary webcam is needed, which makes the system a promising basis for assistive fingerspelling input.