Datasets:
license: cc-by-4.0
language:
- zh
tags:
- audio
- speech
Dataset Card for Dataset Name
This dataset card aims to be a base template for new datasets. It has been generated using this raw template.
Dataset Details
Dataset Description
This dataset contains 50 high-quality audio segments in Chinese, over 3 minutes. The content is based on a transcript discussing the speech about long-term Harvard study on adult development and happiness. These segments were recorded by myself.
Issue Encountered & Solution
I encountered the issue that the inconsistent volume levels across different recording sessions, which could hinder the TTS model's performance. To resolve this, I used Audacity’s "Normalize" to solve it. And the original script contained several Arabic numerals and English acronyms (e.g., "1938"). Directly using these in the rom column would lead to unpredictable phonetic outputs. I resolved this by "normalizing" the text—converting all numbers into full Chinese characters (e.g., "一九三八") in the metadata to ensure the model learns the exact pronunciation intended.