Format Requirements for Audio Datasets

Constraints

• Import from OBS: The size of a single file or compressed package cannot exceed 20 GB. If multiple files are imported, the total file size cannot exceed 20 GB.
• Local import: The size of a single file cannot exceed 1 GB, and the number of files cannot exceed 20.

Format Requirement

Supported file formats: Only audio files with the .mp3, .flac, .wav, .opus, .aac, or .m4a extension and label files with the .jsonl extension are supported. Only UTF-8-encoded JSONL files are supported.

File organization: Supports flexible organization of the dataset structure using multi-level directories (folders).

Annotation file (optional) guidelines: If the dataset contains metadata such as annotation information or audio attributes, it must be implemented by configuring a unified annotation file at a specific path:

1. The annotation file must be located directly in the root directory of the dataset, and its file name must be fixed as annotation.jsonl.
2. Each line in the annotation.jsonl file must be an independent and valid JSON object used to index and describe the attributes of the videos in the dataset.
   • file_name (string, required): Its value must be the relative path of the target audio with respect to the dataset root directory (e.g., audios/speech_01.wav or simply speech_01.wav), and the referenced video file must actually exist at the corresponding path.
   • Custom fields (optional): Aside from file_name, any other key-value pairs are unrestricted. You can customize any service fields according to actual needs.

Data format example

To make the field hierarchy clearer, the single-line JSONL data has been expanded.

{
  "file_name": "audios/mandarin/news_01.wav",
  "text": "Welcome to today's hourly news broadcast.",
  "label": "news_broadcast",
  "language": "zh-CN",
  "audio_properties": {
    "sample_rate": 16000,
    "channels": 1,
    "duration_seconds": 4.8
  }
}