Gunshot Audio Spectrogram Dataset for Binary Classification Using FFT, LogMel, and MFCC Features

This dataset comprises 15,962 labeled audio samples organized into two classes: Gunshot (5,614 instances) and Non-Gunshot (10,348 instances). The data were collected from multiple public repositories, including UrbanSound8k, ESC-50, Gunshot/Gunfire Audio Dataset, Gunshot Audio Dataset, and Gunshot Audio Forensics Dataset. All audio files were pre-processed using the Librosa library: audio durations were standardized to 5 seconds through trimming or zero-padding, and a pre-emphasis filter was applied to enhance high-frequency components. Feature extraction was performed using three established techniques—Fast Fourier Transform (FFT), Mel-Frequency Cepstral Coefficients (MFCC), and Log-Mel Spectrograms—resulting in a spectrogram-based dataset suitable for training and evaluating machine learning models for gunshot detection tasks.本数据集包含15962条带标注的音频样本，划分为两类：枪声（Gunshot）共5614条样本，非枪声（Non-Gunshot）共10348条样本。数据采集自多个公开音频数据集仓库，包括UrbanSound8k、ESC-50、Gunshot/Gunfire Audio Dataset、Gunshot Audio Dataset以及Gunshot Audio Forensics Dataset。所有音频文件均通过Librosa库完成预处理：通过裁剪或零填充将音频时长统一标准化为5秒，并施加预加重滤波器以增强高频分量。本数据集采用三种成熟的特征提取技术——快速傅里叶变换（Fast Fourier Transform，FFT）、梅尔频率倒谱系数（Mel-Frequency Cepstral Coefficients，MFCC）以及对数梅尔频谱图（Log-Mel Spectrograms）——进行特征提取，最终得到基于频谱图的数据集，可用于训练与评估枪声检测任务的机器学习模型。