8Opt/starss22-23-stereo-sample

STARSS22-23 Stereo Subset

A sampled subset (~3GB per split) of the STARSS22-23 dataset (stereo version), packaged for quick experimentation with Sound Event Localization and Detection (SELD) pipelines.

📋 Dataset Description

This dataset is a convenience subset of the STARSS22-23 (Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events), created for rapid prototyping and testing of SELD models on limited compute.

• Source: STARSS22-23 (stereo development split)
• Original Authors: A. Politis, K. Shimada, P. Sudarsanam, S. Adavanne, D. Krause, Y. Koyama, N. Takahashi, S. Takahashi, Y. Mitsufuji, T. Virtanen
• Original Paper: STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events

📊 Dataset Structure

Splits

Split	Description
train	~3GB sampled from dev-train-sony + dev-train-tau
test	~3GB sampled from dev-test-sony + dev-test-tau

Features

Column	Type	Description
file_name	string	Original filename (e.g., fold3_room21_mix013_deg022_start0472)
audio	Audio	Stereo audio waveform, 24kHz
source_type	string	Recording source: sony or tau
annotations	string	Frame-level CSV annotations (see below)
n_frames	int	Total number of annotated frames
n_events	int	Number of unique sound event classes in this file
classes	string	Comma-separated list of event class names

Annotation Format

The annotations column contains a CSV string with frame-level spatial annotations:

Column	Type	Description
frame	int	Frame index (each frame = 100ms)
class	int	Sound event class ID (0–12)
source	int	Source instance ID (tracks same source across frames)
azimuth	int	Horizontal angle in degrees
distance	int	Distance from microphone
onscreen	int	1 = visible on screen, 0 = offscreen

Sound Event Classes (13)

ID	Class
0	Female speech
1	Male speech
2	Clapping
3	Telephone
4	Laughter
5	Domestic sounds
6	Walk / Footsteps
7	Door (open/close)
8	Music
9	Musical instrument
10	Water tap
11	Bell
12	Knock

Audio Details

• Sample Rate: 24,000 Hz
• Channels: Stereo (2 channels)
• Recording Sources: Sony 360RA and TAU Spatial Audio Microphone
• Scenes: Real-world indoor environments

🚀 Usage

from datasets import load_dataset
import pandas as pd
from io import StringIO

ds = load_dataset("your-username/starss22-23-stereo-sample")

# Access a sample
sample = ds["train"][0]
print(sample["file_name"])      # "fold3_room21_mix013_deg022_start0472"
print(sample["classes"])        # "Music,Female speech"
print(sample["audio"])          # {'array': array([...]), 'sampling_rate': 24000}

# Parse annotations back to DataFrame
ann_df = pd.read_csv(StringIO(sample["annotations"]))
print(ann_df.head())
#    frame  class  source  azimuth  distance  onscreen
# 0      0      8       0       24       215         1
# 1      1      8       0       24       215         1

Visualization Example

import matplotlib.pyplot as plt
import numpy as np

sample = ds["train"][0]
ann_df = pd.read_csv(StringIO(sample["annotations"]))

# Plot event timeline
fig, ax = plt.subplots(figsize=(14, 4))
for cls_id in ann_df["class"].unique():
    cls_data = ann_df[ann_df["class"] == cls_id]
    ax.scatter(cls_data["frame"] * 0.1, [cls_id] * len(cls_data), s=2, label=f"Class {cls_id}")
ax.set_xlabel("Time (s)")
ax.set_ylabel("Class ID")
ax.legend()
plt.show()

⚠️ Important Notes

• This is a subset of the full STARSS22-23 dataset, sampled randomly with seed=42
• Only stereo audio is included (not FOA or microphone array formats)
• For official DCASE benchmarking, use the full dataset from Zenodo
• This subset is intended for pipeline testing and prototyping only

📄 Citation

If you use this dataset, please cite the original STARSS23 paper:

@article{politis2023starss23,
  title={STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events},
  author={Politis, Archontis and Shimada, Kazuki and Sudarsanam, Parthasaarathy and Adavanne, Sharath and Krause, Daniel and Koyama, Yuichiro and Takahashi, Naoya and Takahashi, Shusuke and Mitsufuji, Yuki and Virtanen, Tuomas},
  journal={arXiv preprint arXiv:2306.09126},
  year={2023}
}

🏷️ License

This dataset follows the original STARSS23 license (CC BY-NC-SA 4.0). Please refer to the original dataset for full license details.