eduhk-compling/11609227_mahmoodsania

Dataset Overview This dataset contains 50 distinct audio recordings in Urdu, covering everyday conversational topics such as weather, daily routines, and general statements. The audio was self-recorded in a quiet environment using [insert your device, e.g., a dedicated microphone / smartphone voice recorder]. The raw audio was processed into 16-bit PCM WAV format at 44.1kHz and segmented into individual sentence-level clips.

Issues Encountered & Resolution During the preparation process, I encountered two main challenges. Linguistic Issue (Aspiration): Urdu relies heavily on aspirated sounds (like 'bh' in bhi or 'kh' in khushi). Initially, I noticed I was rushing these pronunciations, making them sound flat. To resolve this, I re-recorded specific lines (IDs s006 and s022), focusing on clear enunciation to distinguish between aspirated and non-aspirated consonants. Technical Issue (Segmentation): While using automated silence detection in Audacity, the software occasionally clipped the trailing nasal sounds common in Urdu (like hoon or hain). To fix this, I manually adjusted the label boundaries for every clip to ensure the full decay of the final word was captured without introducing excessive silence.