README.md

metadata

license: cc-by-4.0
language:
  - zh
tags:
  - audio
  - speech
pretty_name: My Audio Dataset

Dataset Card for Assignment 2

Dataset Description (数据集简介)

This is a Chinese speech dataset created for LIN3046 Assignment 2. It contains 50 short sentences recorded by a native Mandarin speaker. The data is segmented into individual WAV files and annotated in the metadata.csv file.

• Total Duration: > 3 minutes
• Language: Chinese (zh)
• Format: WAV (44.1kHz, 16-bit)

• Reflection (反思与挑战)

During the creation of this dataset, I encountered and overcame several technical challenges:

1. Audio Formatting: Initially, I exported files with automatic numbering (e.g., 01-s001.wav), which caused a mismatch with the metadata.csv. I learned to use Audacity's "Label/Track Name" export feature to ensure filenames strictly matched the IDs (e.g., s001.wav).
2. Metadata Encoding: I experienced issues with hidden formatting characters when editing CSV files on macOS. I resolved this by directly editing the metadata.csv on the Hugging Face web interface to ensure strict CSV formatting.
3. Platform Latency: I learned that server-side caching on Hugging Face can sometimes cause temporary RowsPostProcessingError alerts even when the data integrity is correct.

Through this process, I gained a deeper understanding of strict data formatting requirements in computational linguistics.