Newton2676/fireworks-dataset

🎆 Fireworks Audio Dataset

A curated dataset of fireworks sounds, built by merging personal field recordings with filtered clips from FSD50K.
Designed for audio classification, sound event detection, and edge AI deployment.

---

📦 Dataset at a glance

Property	Value
Label	fireworks
Format	WAV PCM 16-bit
Sample rate	44 100 Hz
Channels	Stereo (2)
Clip duration	5 s (fixed)
Splits	train (90 %) · test (10 %)

---

🗂️ Sources

Source	Description	License
Personal recordings	Videos filmed directly at fireworks events — diverse acoustic conditions (distance, venue, crowd)	© Newton2676
FSD50K	FreeSound clips tagged Fireworks (/m/0g6b5 in AudioSet ontology), dev + eval splits	CC-BY (per clip)

---

🔢 Dataset fields

{
    "audio":  Audio(sampling_rate=44100),  # numpy float32 array + sr
    "label":  "fireworks",                 # always "fireworks" in v1
    "source": "personal" | "fsd50k",      # traceability
}

---

🚀 Quick start

from datasets import load_dataset

ds = load_dataset("Newton2676/fireworks-dataset")

sample = ds["train"][0]
audio_array = sample["audio"]["array"]   # numpy float32, shape (N, 2) stéréo
sr          = sample["audio"]["sampling_rate"]  # 44100
source      = sample["source"]           # "personal" or "fsd50k"

---

🔬 Feature extraction example

import librosa
import numpy as np
from datasets import load_dataset

ds = load_dataset("Newton2676/fireworks-dataset")

def extract_features(batch):
    mfccs = []
    for audio in batch["audio"]:
        y  = audio["array"].mean(axis=-1).astype("float32")  # stéréo → mono
        sr = audio["sampling_rate"]
        mf = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=40)
        mfccs.append(mf.mean(axis=1))
    batch["mfcc"] = mfccs
    return batch

ds = ds.map(extract_features, batched=True, batch_size=32)

---

🏗️ Construction pipeline

Personal videos (mp4 / mov / mkv)
         │
         ▼  ffmpeg
   WAV 44.1 kHz stereo
         │
         ▼  pydub
   5 s clips  →  clips/personal/

philgzl/fsd50k  (HuggingFace, streaming)
         │
         ▼  ground_truth CSV filter → /m/0g6b5
   Fireworks fname IDs
         │
         ▼  Opus → WAV 44.1 kHz stereo  (pydub + ffmpeg)
         │
         ▼  pydub
   5 s clips  →  clips/fsd50k_segmented/

         │
         ▼  datasets.Dataset
   Merge → shuffle → 90 / 10 split
         │
         ▼
   Newton2676/fireworks-dataset  🤗

Builder notebook: fireworks_dataset.ipynb

---

📊 Class distribution

This is a single-class dataset (v1). All clips are labelled fireworks.
A multi-class version including additional pyrotechnic sound sub-types (rockets, fountains, aerial bursts) is planned for v2.

---

⚠️ Limitations

• FSD50K clips are crowd-sourced and may contain background noise or mislabels.
• Personal recordings vary in microphone quality and shooting distance.
• No data augmentation has been applied — clips are raw segments.

---

📝 Citation

If you use this dataset, please also cite FSD50K:

@article{fonseca2022fsd50k,
  title   = {{FSD50K}: {An} open dataset of human-labeled sound events},
  author  = {Fonseca, Eduardo and Favory, Xavier and Pons, Jordi and Font, Frederic and Serra, Xavier},
  journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume  = {30},
  pages   = {829--852},
  year    = {2022},
  doi     = {10.1109/TASLP.2021.3133208}
}

---

Built by Newton2676 · part of the STRIX audio ML project series