piyazon/uyghur_music_refined_cleaned

Uyghur Music Refined — Cleaned

1,958 full-length songs, totaling approximately 137.37 hours, with lyrics, music captions, musical key, and tempo. The collection contains Uyghur, Chinese, English, and mixtures of those languages.

Access

Access is gated with manual approval. Sign in to Hugging Face, request access, and wait for the dataset owner to approve your request.

Data

One train split contains exactly five columns:

Column	Type	Description
audio	Audio	Original MP3 bytes embedded in the Parquet files.
lyrics	string	Lyrics in the original script, including section tags.
caption	string	English music description.
key	string	Musical key and mode, for example Bb minor.
bpm	int32	Tempo in beats per minute.

There are no title, artist, original filename, source path, ID, or review-note columns. The audio objects use anonymous numeric filenames. Recordings retain their original MP3 bytes without transcoding or trimming.

Preparation

The source annotations were filtered to remove review-flagged or unclear entries, missing lyrics, missing annotations, and entries containing any language other than Uyghur, Chinese, or English. Instrumentals and unidentified languages were excluded. Every retained song has a matching valid BPM/key analysis and readable audio. Captions and lyrics originate from automated annotations; the owner checked a sample of the cleaned collection before publishing. This is not a claim that every transcription has been manually verified.

The key and BPM come from audio analysis. Any saved review corrections were applied before export. The original collection contained 2,127 MP3s; the initial cleaning retained 2,021. This release contains 1,958 non-excluded records from the reviewed collection.

Uyghur Arabic-script vocabulary filter

This release filters piyazon/uyghur_music_refined at commit c3b743303561753e1bd4d51d2656345c830cb39a using the owner's supplied vocab.json.

• Source songs: 2,021.
• Entire songs excluded: 63.
• Songs retained: 1,958.
• Vocabulary SHA-256: 491e61516353d6144ca81ffa73229ca09e02228366e9945f376cad25c723a8d5.
• Allowed vocabulary characters: !-.:،؛؟ئابتجخدرزسشغفقكلمنوىيپچژڭگھۆۇۈۋېە.

Every lyric character whose Unicode Script_Extensions includes Arabic must be present in the vocabulary. A single out-of-vocabulary Arabic-script character excludes the entire song, including when it appears in multilingual lyrics. This also checks Arabic-script combining marks, elongation characters, digits, and presentation forms. Membership uses the exact original Unicode code point: no normalization, substitutions, or character deletion is performed. Non-Arabic text, such as Chinese, English, section labels, newlines, and neutral punctuation, is retained.

Retained audio, lyrics, captions, key, and BPM match the source release unchanged. The source dataset remains available separately.

Loading

After your access request is approved, authenticate with your Hugging Face account:

from datasets import Audio, load_dataset

dataset = load_dataset("piyazon/uyghur_music_refined_cleaned", split="train", token=True)

# Read the embedded original MP3 without installing an audio decoder:
raw_dataset = dataset.cast_column("audio", Audio(decode=False))
sample = raw_dataset[0]
with open("sample.mp3", "wb") as handle:
    handle.write(sample["audio"]["bytes"])

For automatic decoded audio, install the audio dependencies for your version of datasets.