Datasets:

voidful
/

tkt-song

license: other
task_categories:
  - automatic-speech-recognition
  - audio-text-to-text
tags:
  - hokkien
  - taiwanese
  - lyrics
  - speech
  - audiofolder
pretty_name: TKT Song

TKT Song

Public Taiwanese Hokkien song dataset in Hugging Face AudioFolder layout.

This dataset contains rows where all three resources are available and line-aligned:

song audio converted to 16 kHz mono FLAC
Han-character lyrics in lyrics.han
Hokkien romanization in lyrics.hokkien

Current build row count: 29384.

Schema

audio: Hugging Face Audio feature generated from file_name
source_id: source lyrics id
title: source song title
duration_seconds: YouTube duration when available
lyrics.han: Han-character lyrics
lyrics.hokkien: Hokkien romanization
line_pairs: line-level Han/Hokkien pairs
youtube_url, youtube_id, youtube_title, youtube_channel, youtube_channel_id, youtube_channel_url, youtube_upload_date, youtube_duration_seconds, youtube_view_count, youtube_like_count, youtube_tags: YouTube metadata from yt-dlp
artist_candidates: heuristic artist candidates inferred from YouTube metadata
artist_source: provenance for artist_candidates; heuristic values are not verified ground truth

No YouTube subtitle/caption text is included in this build.

Loading

from datasets import load_dataset, Audio

ds = load_dataset("voidful/tkt-song", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
print(ds[0])