hoyyu1/infant-cry-detection

Dataset Documentation / 数据集说明

Introduction / 简介

This dataset is designed for infant cry detection in noisy household environments. It aims to provide a diverse, representative, and robust collection of audio samples for training and evaluating machine learning models, particularly deep neural networks. 本数据集专为嘈杂家庭环境下的婴儿哭声检测而设计。它旨在提供一个多样化、具代表性且具备鲁棒性的音频样本集合，用于训练和评估机器学习模型（尤其是深度神经网络）。

The experimental dataset was constructed by integrating multiple public datasets to ensure diversity and representativeness. 实验数据集通过整合多个公开数据集构建，以确保样本的多样性和代表性。

The dataset contains approximately 39.6 hours of audio recordings, partitioned into two classes with 5 balanced folds: cry (positive) and non-cry (negative). 该数据集包含约 39.6 小时的音频记录，划分为两类：婴儿哭声（正样本）和非哭声（负样本），共五折供交叉验证（因此不严格区分训练验证测试集）。

The code is at https://github.com/fhfjsd1/ICD_MMSP 代码地址：https://github.com/fhfjsd1/ICD_MMSP

Directory Structure / 目录结构

The dataset is organized into the following directories and files: 数据集组织为以下目录和文件结构：

• audio/: Contains all the audio samples. 包含所有的音频样本文件。
  • positive/: Contains the infant cry (positive) audio samples. 包含婴儿哭声（正样本）音频文件。
  • negative/: Contains the non-cry (negative) audio samples. 包含非哭声（负样本）音频文件。
• metadata/: Contains CSV files with metadata information (such as all28479.csv, noise.csv, noise_val.csv, and rir.csv). 包含存储元数据及标签信息的 CSV 文件（如 all28479.csv、noise.csv、noise_val.csv 以及 rir.csv）。
• rir/: Contains Room Impulse Response (RIR) audio files used for generating reverberation in data augmentation. 包含用于在数据增强中生成环境混响效果的房间冲激响应（RIR）音频文件。

Positive Samples (Cry) / 正样本（婴儿哭声）

The cry data subset consists of audio recordings derived from CryCeleb2023, EnesBabyCries1, and the 2020 iFLYTEK A.I. Developer Competition. 婴儿哭声数据子集包含来自 CryCeleb2023、EnesBabyCries1 以及 2020 科大讯飞 A.I. 开发者大赛的音频记录。

Negative Samples (Non-cry) / 负样本（非哭声）

The non-cry data integrates VoxCeleb, ESC 50, Cat Meowing, and DASEE datasets. 非哭声数据整合了 VoxCeleb、ESC 50、Cat Meowing 和 DASEE 数据集。

It is divided into five subclasses: Speech, Human non-speech, Cat Meows, Household noise, and Silences. 非哭声数据分为五个子类：语音、人类非语音、猫叫声、家庭噪音和静音。

Specific confusable sounds like Cat Meows and various Household noises (domestic appliances, outdoor natural environments, urban traffic) are included. 特别包含了容易混淆的声音（如猫叫声）以及各种家庭噪音（家用电器、户外自然环境、城市交通）。

Data Preprocessing / 数据预处理

Silent segments were removed from the samples as they generally do not contain useful acoustic information. 样本中的静音片段被移除，因为它们通常不包含有用的声学信息。

The remaining segments were split or concatenated to form uniformly sized audio samples. 剩余的片段被分割或拼接，以形成统一长度的音频样本。

All samples were converted and resampled to single-channel 16 kHz WAV PCM format, with a fixed length of 5 seconds. 所有样本均被转换并重采样为单声道 16 kHz WAV PCM 格式，固定长度为 5 秒。

This length is adequate to capture one or multiple cycles of an infant cry. 这个长度足以捕获一个或多个婴儿哭声周期。

One possible processing pipeline used in our experiments for the dataset is as follows: 一种在我们的实验中用于该数据集的可能的数据处理流程如下：

For each processed sample, a Hanning window of size 512 is applied to segment the sample into frames with a hop length of 400.
对于每个处理过的样本，应用大小为 512 的汉宁窗，以 400 的步长将样本分割成帧。

A Mel-scale STFT is computed using 128 triangular filter banks to derive the linear power spectrum.
使用 128 个三角形滤波器组计算梅尔尺度短时傅里叶变换（STFT），以获得线性功率谱。

To achieve regional normalization, the triangular Mel weights are divided by the width of the Mel band, and a logarithmic transformation is subsequently applied to yield log Mel-spectrogram of 128 dimensions.
为了实现区域归一化，将三角形梅尔权重除以梅尔频带的宽度，随后应用对数变换以生成 128 维的对数梅尔频谱图。

The log Mel-spectrogram has been widely used as the input of deep neural networks for many types of audio processing tasks, such as audio clustering, speaker recognition, and audio classification. Hence, it is also used in the study.
对数梅尔频谱图已被广泛用作深度神经网络处理多种音频任务（如音频聚类、说话人识别和音频分类）的输入对象，因此本研究也采用了该特征。

Data Augmentation / 数据增强

Three data augmentation methods (speed perturbation, environmental corruption, time and frequency masking) are executed randomly during each iteration. 在每次迭代中，随机执行三种数据增强方法（速度扰动、环境噪声破坏、时间与频率掩蔽）。

1. Speed Perturbation / 速度扰动

This method involves resampling the audio sample at a random rate similar to the original, resulting in a slightly slower or faster rendition of the signal. 该方法涉及以与原始采样率相近的随机速率对音频样本进行重采样，从而产生稍微变慢或变快的信号效果。

This operation uniformly selects a speed-up factor from 0.8 to 1.2 relative to the original sampling rate and also influences other acoustic features, such as pitch and formant frequencies. 此操作在相对于原始采样率 0.8 到 1.2 的范围内均匀选择一个加速因子，这同时也会影响其他声学特征，例如音高和共振峰频率。

2. Environmental Corruption / 环境噪声破坏

In real-world scenarios, the signals are often contaminated by undesired noises. Consequently, beginning with a clean signal, various types of disruptions are introduced in a controlled manner. 在现实场景中，信号经常会被不想要的噪音所污染。因此，从干净的信号起点，以受控的方式引入各种类型的干扰。

For each pair consisting of a waveform vector $x$ and a noise vector $n$, additive noise is scaled according to a specified signal-to-noise ratio (SNR) and then added to the original waveform, following a defined formula: 对于由波形向量 $x$ 和噪声向量 $n$ 组成的每一对数据，根据指定的信噪比（SNR）对加性噪声进行缩放，然后将其叠加到原始波形上，遵循以下公式：

$$y=x+a\cdot ny\; =\; x\; +\; a\; \backslash cdot\; n$$

where $a$ is a coefficient and defined by: 其中 $a$ 是一个系数，其定义为：

$$a=\sqrt{\frac{\mathrm{\mid }\mathrm{\mid }x\mathrm{\mid }{\mathrm{\mid }}_2^2}{\mathrm{\mid }\mathrm{\mid }n\mathrm{\mid }{\mathrm{\mid }}_2^2}\cdot {10}^{-\frac{SNR}{10}}}a\; =\; \backslash sqrt\{\backslash frac\{||x||\_2^2\}\{||n||\_2^2\}\; \backslash cdot\; 10^\{-\backslash frac\{SNR\}\{10\}\}\}$$

In acoustic environment like rooms, multipath propagation caused by sound reflections, known as reverberation, can significantly degrade the clarity of signals. 在房间等声学环境中，声音反射引起的多径传播（即混响）会显著降低信号的清晰度。

The reverberation effect between the sound source and the receiver is modeled by an impulse response, with which reverberation is introduced via convolution. The room impulse responses employed in this process are obtained from real acoustic measurements. 通过使用冲激响应来模拟声源与接收器之间的混响效果，并通过卷积引入混响。在此过程中使用的房间冲激响应是获取自真实的声学测量数据。

3. Time and Frequency Masking / 时间与频率掩蔽

This technique involves randomly replacing contiguous segments of the original sample with zero values in both the time and frequency domains. 该技术涉及在时间域和频率域中，用零值随机替换原始样本中的连续片段。

Such augmentation encourages the model to rely on information distributed throughout the entire sample rather than particular part. 这种增强方式促使模型依赖分布在整个样本中的信息，而不是特定的局部片段。

Dataset Statistics / 数据集统计

The dataset consists of 13,330 cry (positive) samples and various non-cry (negative) samples. 数据集包含 13,330 个哭声（正）样本和各类非哭声（负）样本。

The number of samples for each specific class is as follows: 各类别的具体样本数量如下：

• Cry / 婴儿哭声: 13,330
• Speech / 语音: 4,963
• Household noise / 家庭噪音: 4,506
• Human non-speech / 人类非语音: 2,076
• Cat Meows / 猫叫声: 1,953
• Silences / 静音: 1,650

Citation / 引用

If you use this dataset in your research, please cite the following paper: (You can find the paper in this dataset) 如果您在研究中使用了该数据集，请引用以下论文：（您可以在这个数据集中找到该论文）

H. Yu and Y. Li, "Infant Cry Detection In Noisy Environment Using Blueprint Separable Convolutions and Time-Frequency Recurrent Neural Network," 2025 IEEE International Workshop on Multimedia Signal Processing (MMSP), Beijing, China, 2025, pp. 1-6, doi: 10.1109/MMSP64401.2025.11324248.