Datasets:
license: cc-by-4.0
language:
- zh
- en
tags:
- audio
- speech
Dataset Description
My dataset consists of 70 short Cantonese audio clips, each paired with a matching Chinese transcription and Jyutping romanisation, designed for basic text‑to‑speech experimentation. I recorded the speech using a mobile phone voice‑recording app, then converted, segmented, and cleaned the audio into individual WAV files, indexed by a metadata.csv file.
Issues Encountered & Solution
I encountered two major difficulties when preparing the data. First, numerous brief pauses were not recognized as quiet by Audacity's automatic segmentation, which resulted in misaligned clips. I resolved this by tweaking the threshold level and minimum silence duration till the borders matched my sentences more accurately.Second, the recordings had constant background noise from surrounding traffic and air conditioning, which could degrade TTS quality. I re-recorded in a calmer setting, used mild noise reduction, and trimmed leading and trailing silence to lessen this. The outcome was clearer, more reliable audio clips.