laupeng1989/audio-reconstruct-universal-testset-v2

Audio Reconstruction Universal Test Set (v2)

Universal evaluation set for audio reconstruction / neural-codec resynthesis. It covers three domains — speech, vocals and sound effects — with 300 clips each (900 total, 1.82 hours), all mono 24 kHz WAV, 1.0–30.0 s long (mean 7.3 s).

This is v2, the official release. It is the set used to evaluate RFWave vocoders, including laupeng1989/rfwave-encodec-24k.

Layout

speech/        vocals/        sound_effect/
  <subset>/<clip>.wav
readme.pdf     # original release notes
metadata.csv   # file_name, category, subset, duration_sec, sample_rate, channels

How to load

1. Download the whole set

from huggingface_hub import snapshot_download

root = snapshot_download(
    "laupeng1989/audio-reconstruct-universal-testset-v2",
    repo_type="dataset",
)

2. Iterate the audio

import glob, os
import soundfile as sf

files = sorted(glob.glob(os.path.join(root, "*", "*", "*.wav")))   # 900 files
print(len(files))

wav, sr = sf.read(files[0])      # sr == 24000, mono, float64 in [-1, 1]

The category is the first path component (speech / vocals / sound_effect), the subset the second one (libritts_test, opencpop, ...).

3. Use the metadata table

import os
import pandas as pd

meta = pd.read_csv(os.path.join(root, "metadata.csv"))
print(meta.head())
print(meta.groupby("category")["duration_sec"].agg(["count", "sum"]))

Column	Meaning
file_name	path relative to the dataset root
category	speech, vocals or sound_effect
subset	source corpus, e.g. libritts_test
duration_sec	duration in seconds
sample_rate	always 24000
channels	always 1

4. Fetch a single clip

from huggingface_hub import hf_hub_download

p = hf_hub_download(
    "laupeng1989/audio-reconstruct-universal-testset-v2",
    "speech/libritts_test/LibriTTS_test-clean_1580_141083_1580_141083_000070_000000.wav",
    repo_type="dataset",
)

Do not use load_dataset("audiofolder", data_dir=...) here. The *_test subset directory names are picked up by the builder's split inference, so it silently loads only 410 of the 900 files as a test split. Use snapshot_download as shown above.

Composition

Category	Subset	Clips
speech	aishell3_test	50
speech	cml_tts_test	50
speech	expresso_test	50
speech	hifi_tts_test	50
speech	jvs	50
speech	libritts_test	50
vocals	CSD	20
vocals	ccmusic_demo_vocals	20
vocals	chinese_opera_polyphonic	60
vocals	jvs_music	60
vocals	musdb_test_vocal	60
vocals	opencpop	20
vocals	ravdess	60
sound_effect	ccmusic_demo_music	30
sound_effect	chinese_opera_polyphonic	20
sound_effect	esc_50	150
sound_effect	musdb_test_accompaniment	40
sound_effect	musdb_test_mix	20
sound_effect	openmic_2018_test	40
Total	19 subsets	900

Notes

• All files are already 24 kHz mono; no resampling needed for 24 kHz models.
• macOS metadata files (.DS_Store, ._*) were removed from this mirror.
• metadata.csv was generated with soundfile from the audio headers.

Licensing

The clips originate from third-party corpora that keep their own licenses — some are non-commercial (e.g. ESC-50, Expresso, MUSDB18). This mirror is provided for research and evaluation only; please check and follow each source corpus's license before any other use.