Real-Time Azerbaijani Sign Language Sentence Recognition: Comparative Evaluation of Cosine Retrieval and Seq2Seq Approaches
Year: 2026
Type: conference-paper
Abstract: Deaf and Hard-of-Hearing people face serious difficulties when communicating with the hearing world. This problem becomes even more serious for low-resource sign languages such as Azerbaijani Sign Language, because there is no large dataset or pretrained model available for this language. For that reason, this work focuses on building and evaluating sentence-level Azerbaijani Sign Language recognition methods that can work on a normal CPU without requiring GPU hardware. To begin with, two main approaches were implemented and tested. The first approach is a training-free cosine retrieval system, which uses MediaPipe hand landmarks. The second approach is a Seq2Seq neural network baseline. The cosine retrieval system achieved 30.7% Top-1 accuracy and 45.0% Top-3 accuracy without any training. On the other hand, the Seq2Seq model achieved 35.5% Top-1 accuracy and a BLEU-4 score of 31.43. However, the Seq2Seq model also showed a limitation called mode collapse because the dataset was small. It produced the same sentence for 36.7% of predictions. Even though this is a weakness, the model’s accuracy and BLEU-4 score still show that its predictions were connected to the visual input and were not random. As a result, both approaches create the first reproducible sentence-level reference point for Azerbaijani Sign Language recognition.
Institution: ADA University
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202076290