Dataset Card for CrowdioSet

Dataset Summary

CrowdioSet is an audience-noise dataset for live music source separation, comprising two components:

• Ambiences & Events: 4,819 real recordings sourced from Freesound (queries: crowd, audience, cheering, applause, chatter, protest), manually classified into two categories: ambiences (room tone, background chatter) and events (applause, cheering, shouts). Provided as 44.1 kHz, 16-bit stereo WAV files (105.12 hours total).
• Synthetic Sing-Alongs: A preliminary generative pipeline producing synthetic sing-along stems for every vocals stem in MUSDB18HQ and MOISESDB datasets, combining the Antares AVOX Choir effect with zero-shot singing voice conversion (HQ-SVC). For validation and test, an additional already pre-mixed audience track (ambiences + events already combined) is provided per song.

Supported Tasks

• Live music source separation / audio-to-audio: the ambience, event, and sing-along stems are intended to be mixed with clean studio stems (e.g. MUSDB18HQ, MOISESDB) to synthesize realistic live-recording training data for source separation models.

Languages

Not applicable — the dataset consists of non-linguistic audio (crowd noise, cheering, applause, singing).

Dataset Structure

Data Instances

Each ambience/event instance is a single WAV file. Each sing-along instance is a folder named after the corresponding MUSDB18HQ/MOISESDB track, containing:

• singalong.wav — the synthetic sing-along stem (train), or
• singalong.wav + audience.wav (valid/test) — sing-along stem plus the pre-mixed ambience+event audience track for reproducibility.

Data Fields

• audio: 44.1 kHz, 16-bit stereo WAV file.
• Per-file license and attribution metadata (source, license type, uploader, Freesound ID) is provided separately in crowdioset_metadata.csv for every ambience/event recording.

Data Splits

Split	Ambiences	Events	Sing-alongs
train	2,345	2,346	320 (MUSDB18HQ train + MOISESDB)
valid	14	14	14
test	50	50	50

Splits follow the MUSDB18HQ convention (50 tracks reserved for test, 14 for validation). Sing-alongs are provided for every MOISESDB song and for the full MUSDB18HQ training set. In valid and test, ambiences and events are already pre-mixed per track into audience.wav.

Loading with PyTorch

CrowdiosetAmb / CrowdiosetEve sample fixed-length crops from train/ambiences and train/events:

import os
import torch
import torchaudio
from torch.utils.data import Dataset
import itertools

def _load_stereo(path, sample_rate):
    wav, sr = torchaudio.load(path)
    if sr != sample_rate:
        wav = torchaudio.functional.resample(wav, sr, sample_rate)
    return wav.repeat(2, 1) if wav.shape[0] == 1 else wav[:2]


class CrowdiosetAmb(Dataset):
    def __init__(self, root, duration, sample_rate=44100):
        self.dir = os.path.join(root, 'train', 'ambiences')
        self.files = sorted(f for f in os.listdir(self.dir) if f.endswith('.wav'))
        self.duration, self.sample_rate = duration, sample_rate

    def __len__(self):
        return len(self.files)

    def __getitem__(self, idx):
        wav = _load_stereo(os.path.join(self.dir, self.files[idx]), self.sample_rate)
        while wav.shape[1] < self.duration:
            wav = wav.repeat(1, 2)
        start = torch.randint(0, wav.shape[1] - self.duration + 1, (1,)).item()
        return wav[:, start:start + self.duration]


class CrowdiosetEve(Dataset):
    def __init__(self, root, duration, sample_rate=44100):
        self.dir = os.path.join(root, 'train', 'events')
        self.files = sorted(f for f in os.listdir(self.dir) if f.endswith('.wav'))
        self.duration, self.sample_rate = duration, sample_rate

    def __len__(self):
        return len(self.files)

    def __getitem__(self, idx):
        wav = _load_stereo(os.path.join(self.dir, self.files[idx]), self.sample_rate)[:, :self.duration]
        out = torch.zeros(2, self.duration)
        start = torch.randint(0, self.duration - wav.shape[1] + 1, (1,)).item()
        out[:, start:start + wav.shape[1]] = wav
        return out

Training

# sources_dataset = your musdb dataset class
# sources_loader = DataLoader(srcs_dataset, batch_size, shuffle=True, drop_last=True)
ambs_dataset = CrowdiosetAmb(root, duration) #duration in samples
eves_dataset = CrowdiosetEves(root, duration)
ambs_loader = DataLoader(ambs_dataset, batch_size, shuffle=True, drop_last=True)
eves_loader = DataLoader(eves_dataset, batch_size, shuffle=True, drop_last=True)

singalong.wav is not loaded through a class — load it as the 5th source (sources[:, 4]), matched by track name, the same way you already load the 4 MUSDB18HQ/MOISESDB stems.

Generate the audience mixture on the fly in the training loop:

p_sing = p_amb = p_eve = 0.5

for sources, ambs, eves in zip(sources_loader, itertools.cycle(ambs_loader), itertools.cycle(eves_loader)): 
    audience = torch.zeros_like(ambs)
    for i in range(len(audience)):
        if torch.rand(1) < p_sing:
            audience[i] += 3 * torch.rand(1).cuda() * sources[i, 4]
        if torch.rand(1) < p_amb:
            audience[i] += 0.5 * torch.rand(1).cuda() * ambs[i]
        if torch.rand(1) < p_eve:
            audience[i] += 0.5 * torch.rand(1).cuda() * eves[i]
        # normalize if needed
        if torch.max(torch.abs(audience[i])) > 1.:
            audience[i] = audience[i] / torch.max(torch.abs(audience[i]))
    sources[:, 4] = audience
    
    mix = sources.sum(dim=1)

Validation

valid/test tracks ship a pre-mixed, deterministic audience.wav (ambiences + events already combined, under valid/<track>/ or test/<track>/) — skip the random mixing above and drop it straight into the source slot:

sources[:, 4] = audience  # audience.wav for this track — deterministic, no random mixing
mix = sources.sum(dim=1)

Dataset Creation

Source Data

Ambience and event recordings were collected from Freesound under CC0, CC-BY 3.0/4.0, CC-BY-NC 3.0/4.0, and Sampling+ 1.0 licenses, then manually classified into ambiences vs events. Sing-along stems were synthetically generated from MUSDB18HQ and MOISESDB vocal stems using the Antares AVOX Choir effect combined with zero-shot singing voice conversion (HQ-SVC).

Annotations

Category labels (ambience vs. event) were assigned manually. No other annotations are provided.

Considerations for Using the Data

Licensing Information

91.98% of the ambience/event files are usable in commercial applications (58.79% CC0, 33.19% CC-BY); the remaining 8.02% are CC-BY-NC (non-commercial only). A small subset is licensed under Creative Commons Sampling+ 1.0, which does not map to a standard SPDX identifier.

Per-file license and attribution metadata for every individual audio file is provided in crowdioset_metadata.csv. Consult this file before using or redistributing any specific recording, as licensing terms vary file by file — the aggregate license field above covers the range of licenses present but does not guarantee any single file is unrestricted.

Sing-along stems inherit the licensing terms of their source dataset (MUSDB18HQ / MOISESDB) plus any restrictions from the Antares AVOX Choir effect and the voice conversion model used to generate them.

Social Impact and Biases

Freesound-sourced recordings may contain identifiable ambient speech from crowds; no attempt was made to anonymize incidental background voices.

Additional Information

Dataset Curators

Enric Gusó

Licensing Information

See Licensing Information above.

Citation Information

Guso, E., Serra, X. (2026). CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation. Proceedings of the 27th ISMIR Conference, 2026 nov 8-12, Abu Dhabi, UAE

Contributions

Audio examples: https://enricguso.github.io/crowdioset_parirset Paper: https://arxiv.org/pdf/2607.27828