Balti-Tamko: The first Balti spoken words dataset

Balti Speech Dataset This dataset comprises 10,394 high-quality voice recordings capturing approximately 500 distinct isolated Balti words, primarily nouns. The recordings were collected from 39 native Balti speakers representing all five geographical regions of Baltistan in the Gilgit-Baltistan territory of Pakistan. Each audio sample is accompanied by its corresponding text transcription, ensuring robust training capabilities for Automatic Speech Recognition (ASR) systems. ASR Training Experiments The dataset has been rigorously evaluated through two primary training methodologies: Training from Scratch Implemented using GMM-HMM and TDNN architectures via the Kaldi toolkit. Fine-Tuning Pre-Trained Models Leveraged OpenAI’s Whisper models, fully fine-tuned to optimize performance for Balti speech recognition. Performance & Applications The experiments demonstrate high accuracy in isolated Balti word recognition, making this dataset a valuable resource for: Low-resource language ASR development Linguistic research on the Balti language Preservation and digitization of regional dialects This dataset bridges a critical gap in speech technology for underrepresented languages, offering a reliable foundation for future research and applications in Balti speech processing.