eduhk-compling/11557577_PoonYanKei

Dataset Description

This dataset contains 50 Cantonese daily sentences for Text-to-Speech (TTS) model training. The audio totals approximately 4 minutes of clean, segmented recordings. The sentences feature authentic colloquial expressions, longer sentence structures, and are accompanied by accurate Jyutping romanization in the metadata for pronunciation guidance.

Collection Process: The Cantonese sentences were recorded in a quiet environment using a USB condenser microphone. All sentences were recorded in one continuous take, then segmented into individual files. The content includes common daily conversations and natural spoken patterns representative of contemporary Hong Kong Cantonese.

Issues Encountered & Solution

I encountered two main issues: First, segmenting sentences proved challenging due to the language's tonal nature and rapid speech patterns. The automatic silence detection often cut mid-sentence during connected speech. I resolved this by manually adjusting segment boundaries in Audacity, ensuring natural breaks between phrases while maintaining the 3-10 second duration target. Second, ensuring accurate Jyutping transcription required careful verification, particularly for colloquial contractions and sentence-final particles. I cross-referenced multiple pronunciation resources and listened repeatedly to each audio segment, making adjustments to the Jyutping where my natural speech deviated from standard romanization. These processes resulted in accurately segmented and transcribed audio files suitable for TTS training.