sunbv56/song_dataset_chunked

Vietnamese Songs Word-Level Timestamp Dataset (Chunked)

This dataset contains word-level timestamp information for Vietnamese songs, specifically pre-chunked into segments up to 30 seconds for use in training or fine-tuning speech recognition (ASR) systems like Whisper.

Dataset Summary

The song_dataset_chunked provides high-quality Vietnamese song data, properly segmented into optimal ~30-second sequences.

Duration Insights:

• Train split (train_chunked.jsonl): ~ 230.62 hours of audio

This dataset is optimally designed for tasks such as:

• Training and evaluating automatic speech recognition (ASR) models on music.
• Lyrics synchronization (Lyrics Alignment / Karaoke generation).
• Natural language processing (NLP) analysis on song lyrics.

Fields

• chunk_id (string): A unique identifier for the chunk, formatted as {song_id}_{chunk_index:03d}.
• song_id (string): An 8-character string identifier for the original song.
• title (string): The name of the song.
• artist (string): The performer(s) of the song.
• album (string): The album the song belongs to.
• streaming_url (string): URL to stream the audio (MP3/M4A). URL contains temporary security tokens.
• chunk_start_ms (integer): Start time of the chunk in milliseconds (relative to the whole audio file).
• chunk_end_ms (integer): End time of the chunk in milliseconds (relative to the whole audio file).
• chunk_lyrics (string): Text of the lyrics contained in this chunk. Sentences are separated by \n.
• chunk_word_timestamps (list of lists of dicts): The word-level alignments. The top-level list represents sentences/lines, the inner list represents words within the sentence, and the dict contains start, end, and word keys.

Example Single Record

{
  "chunk_id": "ZZ8CC7AZ_001",
  "song_id": "ZZ8CC7AZ",
  "title": "Đợi Chờ Bóng Xuân",
  "artist": "Eric Toàn Nguyễn",
  "album": "Trách Duyên Bẽ Bàng",
  "streaming_url": "https://a128-z3.zmdcdn.me/...",
  "chunk_start_ms": 39340,
  "chunk_end_ms": 66960,
  "chunk_lyrics": "Lặng nhìn sương rơi\nMấy mùa mưa qua",
  "chunk_word_timestamps": [
    [
      {"start": 39340, "end": 39850, "word": "Lặng"},
      {"start": 39850, "end": 40340, "word": "nhìn"},
      {"start": 40340, "end": 40850, "word": "sương"},
      {"start": 40850, "end": 43850, "word": "rơi"}
    ],
    [
      {"start": 43860, "end": 44370, "word": "Mấy"},
      {"start": 44370, "end": 44860, "word": "mùa"},
      {"start": 44860, "end": 45870, "word": "mưa"},
      {"start": 45870, "end": 48390, "word": "qua"}
    ]
  ]
}

How the chunks were created

The data was reconstructed using a chunking heuristic based on capital letters and original lyrical line breaks. It traverses a 2D array representation of the timestamps (sentences x words) and safely slices chunks to ensure they are <=30 seconds. This optimizes the data to be used in standard Whisper sequence-to-sequence model fine-tuning.

How to use

from datasets import load_dataset

dataset = load_dataset("sunbv56/song_dataset_chunked")

# Access the first train chunk
print(dataset['train'][0]['chunk_lyrics'])