EarResp-ANS : Audio-Based On-Device Respiration Rate Estimation on Earphones with Adaptive Noise Suppression

TL;DR

This work introduces EarResp-ANS, a fully on-device pipeline for respiration-rate estimation using in-ear and outer-ear microphones on commodity ANC earphones. It combines a delayed LMS-based adaptive noise suppression stage with a lightweight spectral-feature RR estimator that leverages harmonic cues, and fuses binaural estimates with outlier rejection to achieve robust performance under realistic noise up to 80 dB SPL. Across 18 participants and diverse acoustic conditions, the method attains a mean absolute error of $0.84$ CPM, dropping to $0.47$ CPM after channel-discrepancy-based outlier filtering, and up to $0.14$ CPM within a robust $3\sigma$ interval, while consuming less than $2\%$ of on-device compute and requiring no raw audio to leave the device. This demonstrates the feasibility of privacy-preserving, real-time respiratory monitoring on everyday earphones and opens avenues for richer on-device respiratory analytics and multi-modal fusion in wearables.

Abstract

Respiratory rate (RR) is a key vital sign for clinical assessment and mental well-being, yet it is rarely monitored in everyday life due to the lack of unobtrusive sensing technologies. In-ear audio sensing is promising due to its high social acceptance and the amplification of physiological sounds caused by the occlusion effect; however, existing approaches often fail under real-world noise or rely on computationally expensive models. We present EarResp-ANS, the first system enabling fully on-device, real-time RR estimation on commercial earphones. The system employs LMS-based adaptive noise suppression (ANS) to attenuate ambient noise while preserving respiration-related acoustic components, without requiring neural networks or audio streaming, thereby explicitly addressing the energy and privacy constraints of wearable devices. We evaluate EarResp-ANS in a study with 18 participants under realistic acoustic conditions, including music, cafeteria noise, and white noise up to 80 dB SPL. EarResp-ANS achieves robust performance with a global MAE of 0.84 CPM , reduced to 0.47 CPM via automatic outlier rejection, while operating with less than 2% processor load directly on the earphone.

Figures (11)

• Figure 1: System pipeline of EarResp-ANS on OpenEarable 2.0. ANS denoises the IEM signal on the DSP using a delayed LMS filter. RR estimation (STFT-based features and peak selection) runs on the MCU, with per-ear estimates transmitted via BLE for fusion.
• Figure 2: IEM signals and respiration belt ground-truth data before (A) and after (B) adaptive noise suppression, recorded during the music sound condition. The figure shows the raw time-domain signals together with their corresponding spectrograms. The inspiration and expiration segments vanish under the background noise in figure (A) and are recovered in figure (B) using ANS. Correlation with respiration belt signal becomes visible in both the raw audio signal and the spectrogram.
• Figure 3: This figure shows the feature signal $c(t)$ compared to the denoised IEM signal. The IEM signal has been scaled for appropriate display. The feature signal contains local maxima for each inspiration and expiration.
• Figure 4: Experimental setup, recording protocol, and recording conditions. Participants wore binaural OpenEarable earphones and a respiBAN respiration belt while seated. Audio was played via stereo speakers (approx. 50 cm, head level). Each participant completed six recording sessions in randomized order. For each session, breathing modulation (none or exercise-induced via 30 s jumping jacks) and one acoustic condition were assigned. Acoustic conditions included a low-noise baseline (<35 dB SPL), white noise (50/65/80 dB SPL), cafeteria noise (approx. 80 dB SPL), and music (approx. 80 dB SPL); each condition was recorded once per participant.
• Figure 5: (A) Violin plot showing the distribution of ground-truth respiration rates for each participant. Participants are ordered post hoc according to their mean rate across all recordings. (B) Histogram showing the distribution of ground-truth respiration rates with and without prior physical activity.