UrduSpeakXLSR: Multilingual Model for Urdu Speech Recognition

Speech recognition, a disruptive technology, has revolutionized human-machine interaction. While numerous Automatic Speech Recognition (ASR) models are publicly available via HuggingFace, the majority cater to English language. For Urdu, however, models are scarce or closed-source, with open-sourced ones often lack the robustness. Our research addresses this scarcity, focusing on the challenges posed by dialects, slangs, and accents. In this work, we introduce an innovative ASR model leveraging the pretrained XLS-R model based on Wav2Vec2.0 architecture, trained on CommonVoice corpus V11. Our approach outperforms other deep learning-based techniques both qualitatively and quantitatively, offering promising results for an under-resourced language.