Respiratory Rate Dataset - BreathMY_v2

An open-source audio dataset for respiratory rate (RR) estimation research. About The Dataset The dataset has been generated by selecting 150 recordings from the BreathMY dataset, thereby extending it. This novel dataset comprises three distinct repositories of respiratory signals, specifically developed for the rigorous design and evaluation of RR estimation methodologies, covering both noise free and noise-affected scenarios. The constituent datasets include: DR: dataset of noise-free respiratory signals with controlled RR variability, generated through data augmentation techniques. DSN: dataset with additive white Gaussian noise (AWGN) at different signal-to-noise ratio (SNR) levels (0, -6 and -20 dB). DN: dataset with environmental acoustic noise at different SNR (0, -6 and -20 dB). File Format Each audio file is encoded in .wav format and follows the naming convention: DR_9RR_2023_03_15_4_A.wav Where: DR → Identifies the source dataset (e.g., DR). 9RR → Ground truth RR of the signal, expressed in breaths per minute (bpm). Here, RR = 9 bpm. 2023_03_15 → Date of creation or processing of the signal (YYYY_MM_DD). 4 → Patient index recorded on that specific date. Here, it corresponds to the fourth patient recorded on 2023-03-15. A → Alphabetical character used to differentiate files and prevent overwriting when applying data augmentation DR: Dataset with Noise-Free Respiratory Signals The DR dataset consists of 2,550 respiratory signals, each 60 seconds in duration. It was built using two data augmentation techniques applied to the pre-existing dataset: Time Stretching (TS): transforms signals with fast RR into signals with slower RR. It preserves the original waveform but increases the duration of each respiratory cycle. Cycle Replication (CR): transforms signals with slow RR into signals with faster RR. It compresses the duration of each respiratory phase (inhalation and exhalation) and replicates the cycles to complete the recording interval. DSN: Dataset with Additive White Gaussian Noise To evaluate the robustness of the methods under controlled noise conditions, the DSN dataset is generated from DR by adding AWGN to each signal in the DR dataset at three different SNR levels: DSN0 → SNR = 0 dB DSN-6 → SNR = -6 dB DSN-20 → SNR = -20 dB AWGN uniformly affects all frequency components and represents a standard yet demanding scenario for assessing the performance degradation of algorithms. Each subset preserves the same number of signals as DR, resulting in a total of 7,650 signals — that is, 2,550 respiratory signals mixed with AWGN noise for each SNR level. DN: Dataset with Environmental Acoustic Noise The DN dataset simulates real-world environments by adding environmental acoustic noise to each signal from DR at three SNR levels, containing a total of 7,650 signals — that is, 2,550 respiratory signals mixed with environmental acoustic noise for each SNR level: DN0 → SNR = 0 dB DN-6 → SNR = -6 dB DN-20 → SNR = -20 dB The noises used originate from two types of environments: Clinical (Soundsnap): medical equipment, conversations, footsteps in corridors and other typical hospital sounds. Urban (DCASE): ten different acoustic scenes, including streets, public transport, indoor spaces, etc. For each mixture: The type of noise is assigned randomly. A random 60-second segment of noise is extracted to ensure variability. A balanced representation of all noise types is maintained. Noise segments used for training and testing are mutually exclusive, ensuring that no segment appears in both sets and thereby reducing the risk of overfitting. License Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Copyright (c) 2026 QHPC & SP Research Lab This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. You are free to: - Share — copy and redistribute the material in any medium or format - Adapt — remix, transform, and build upon the material Under the following terms: - Attribution — You must give appropriate credit. - NonCommercial — You may not use the material for commercial purposes. Full license text: https://creativecommons.org/licenses/by-nc/4.0/ Contact Alejandro Antonio Salvador Navarro ([e-mail])University of Jaén Francisco Jesús Cañadas Quesada ([e-mail])University of Jaén Juan de la Torre Cruz ([e-mail])University of Jaén Citing When using this dataset, please cite the following publication: Salvador-Navarro, A., De La Torre-Cruz, J., Muñoz-Montoro, A. J., Ranilla-Cortina, S., Carabias-Orti, J. J., Cruz-Molina, J. M., & Cañadas-Quesada, F. J. (2026). Respiratory rate estimation from breath sounds based on deep learning*. Biomedical Signal Processing and Control, 119, 109905. Funding This work was supported in part by the PID2023-146520OB-{C21,C22} funded by MICIU/AEI/10.13039/501100011033 and, as appropriate, by “ERDF A way of making Europe”, by “ERDF/EU”, by the “European Union” or by the “European Union NextGenerationEU/PRTR” and in part by “REPERTORIUM” Project under Grant Agreement 101095065. Horizon Europe. Cluster II. Culture, Creativity and Inclusive society. Call HORIZON-CL2-2022-HERITAGE-01-02.