pretty_name: KunSet
license: other
license_name: kunset-research-only
language:
- zh
tags:
- kunqu
- opera
- singing-voice
- lyrics-alignment
- forced-alignment
- rhyme
size_categories:
- 10K<n<100K
configs:
- config_name: recordings
default: true
data_files:
- split: train
path: splits/recordings/train.jsonl
- split: dev
path: splits/recordings/dev.jsonl
- split: test
path: splits/recordings/test.jsonl
extra_gated_prompt: >-
KunSet contains audio (full mixes and separated vocals) from recordings of the
Kunqu Opera Art Canon and their annotations. It is provided for non-commercial
academic research only. Do not redistribute the audio or use it for commercial
purposes.
extra_gated_fields:
Name: text
Affiliation: text
Intended use: text
I will use KunSet for non-commercial academic research only and will not redistribute the audio: checkbox
KunSet
Corpus and trained aligner for KunSet: A Corpus and Benchmark for Lyrics Alignment and Rhyme Retrieval in Kunqu Opera. Code, the Yunxue Lizhu rhyme database and the result records are at github.com/Kunyuan-Project/KunSet.
281 recordings: 155.2 h of media, 106.3 h annotated, 482,164 syllable-level timestamps.
Contents
| Path | Contents |
|---|---|
splits/recordings/{train,dev,test}.jsonl |
The table of the dataset viewer: one row per recording with its split, duration, sentence and syllable counts and the paths of its vocal, mix and annotations. The split is by recording; the experiments read it from corpus/corpus.json |
corpus/ |
Frozen evaluation corpus: corpus.json (recordings, 251/20/10 Train/Dev/Test split) and sentences.jsonl (sentence and syllable timestamps) |
audio/vocal/<recording_id>.wav |
Separated vocals, 281 files; the input to every experiment |
audio/mix/<recording_id>.wav |
Full mixes (44.1 kHz stereo 16-bit), 281 files; audio/mix/manifest.json lists their SHA-256 |
kunset/annotations/<recording_id>.json |
KunSet annotations: sentences (subtitleLines[]) and syllables (syllableAnnotations[]) with timestamps, delivery mode, roles and Yunxue Lizhu links (phonology: rhyme group, tone, yin–yang, voicing, wuyin, fanqie) |
kunset/lexicon/ |
Yunxue Lizhu entries referenced by the links |
kunset/manifest.json, kunset/coverage.json |
Provenance and per-attribute coverage |
annotations/annotations/<corpus_id>.json |
The same annotations without dictionary links, including sentences of unknown delivery mode; inputs of the experiments |
annotations/blind_pool/ |
Blind A/B annotations of the 10 Test recordings |
annotations/release.json |
Original to released SHA-256 of each annotation file |
annotations/manifest.json |
Corpus-statistics manifest (Table 1) |
annotations/blind-check-latest10.json |
Blind-agreement selection record (Table 2) |
checkpoints/kunset-aligner.ckpt |
Trained SOFA aligner, 251 training recordings (seed 2027, step 8000, lowest Dev MAE); weights and hyperparameters only, without optimizer and trainer state |
Paths inside corpus/ are relative to data_root in corpus.json. Every vocal
and annotation file is listed there with its SHA-256.
The vocals were separated from the mixes with Kim's Mel-Band RoFormer vocal model
(vocals_mel_band_roformer.ckpt) in audio-separator 0.44.2;
SOFA/scripts/separate_vocals.sh in the GitHub repository reproduces the step.
Re-separated vocals are close to, but not byte-identical with, the released ones.
Usage
hf download Kunyuan-Project/KunSet --repo-type dataset --local-dir data
The recording table loads without the audio (after accepting the terms and logging in):
from datasets import load_dataset
recordings = load_dataset("Kunyuan-Project/KunSet") # train / dev / test
The experiments read corpus/ and annotations/ directly. Run them from a
checkout of the GitHub repository, with this dataset as its data/ directory. With the anchor2_v3 decoder, the aligner reaches a Test MAE of
194.03 ms; Table 3 of the paper reports the mean over three training seeds.
Terms
For non-commercial academic research only. The recordings remain the property of their rights holders.