Sensing Bone-Conducted Speech with Earbuds
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Sensing Bone-Conducted Speech with Earbuds".
Dev: Here's the gist of the paper. Clear capture of the wearer’s own voice (OV) is essential when using earbuds for mobile communication<ref:2609.02165#pg2> Bone-conducted (BC) speech, which can be sensed as vibrations of the earbud housing, can be used to improve OV capture<ref:2609.02165#pg4>.
Rosa: This study investigates both bandwidth and spatial characteristics of OV-induced earbud vibrations based on measurements with two earbud models<ref:2609.02165#pg2>.
Dev: Spectrally, results indicate that OV-induced earbud vibrations exhibit a low-pass characteristic, with a steep roll-off of −93 dB per decade above 400 Hz<ref:2609.02165#pg2>.
Rosa: Spatially, results indicate that the earbuds mainly vibrate in and out of the ear canal entrance, with high consistency between subjects and fits<ref:2609.02165#pg2>.
Paper summary: Rosa: So, this paper, "Sensing Bone-Conducted Speech with Earbuds," it's looking at how we can use those vibrations from bone conduction technology to capture someone's voice when they're using earbuds for mobile communication. The core idea is that the earbud housing itself is vibrating when someone speaks, and we can sense that vibration as a way to get better voice capture than just listening to the air.
Dev: Yeah, it sets out to look at two main things with these earbud vibrations: first the bandwidth, which is how wide a range of frequencies they capture, and second the spatial characteristics—basically where those vibrations are coming from in space. It’s important because you need to know that stuff before you even think about choosing a sensor or figuring out where to place it <ref:2609.02165#pg4>.
Taro: I wonder what that means for real-world use. If we can figure out the bandwidth and the spatial pattern, does it help us design systems that actually work when things get messy? Like, when someone is walking or driving and their environment is noisy <ref:2609.02165#pg4>.
Rosa: Exactly. The paper investigates these characteristics based on measurements from two different earbud models to see what they find about the vibrations induced by the wearer's own voice <ref:2609.02165#pg2>. It’s trying to nail down what these signals actually look like when you measure them.
Dev: From a spectral point of view, they found that these vibrations have a low-pass characteristic, meaning most of the signal power is below a certain frequency, and there's a steep roll-off of minus ninety-three decibels per decade above four hundred hertz <ref:2609.02165#pg2>. That means if you want to sense these vibrations above that four hundred hertz mark, you're going to need a sensor that has a really low noise floor, because the signal drops off so fast <ref:2609.02165#pg5>.
Taro: A steep roll-off like minus ninety-three decibels per decade is quite specific. What does that imply about the frequencies we're actually getting from speech? Is it just low frequencies?
Paper summary: Rosa: It suggests that the most powerful vibrations, the ones you want to focus on, are concentrated below four hundred hertz <ref:2609.02165#pg5>. The paper also found that this amplitude stays roughly constant around five hundred thirty micrograms per pico hertz between one hundred hertz and four hundred hertz <ref:2609.02165#pg5>.
Dev: That's a concrete number for the power level in that band, which is helpful for designing the system's sensitivity <ref:2609.02165#pg5>. But what about where these vibrations are coming from spatially? Where are they actually happening in relation to the ear canal?
Taro: That’s where I get interested. The paper claims that spatially, the earbuds mainly vibrate in and out of the ear canal entrance <ref:2609.02165#pg9>. It also mentions that there's high consistency between different subjects and different fits when measuring this spatial pattern <ref:2609.02165#pg9>.
Rosa: That consistency is a big deal because it suggests that the direction of vibration is pretty predictable, regardless of who you are or how the earbud fits on someone’s ear <ref:2609.02165#pg9>. And they found that this pattern in one direction means you might only need a single-axis accelerometer to capture those high-power components <ref:2609.02165#pg9>.
Dev: That leads right into the next part of their work, which looks at how effective a single-axis sensor actually is by simulating the three-axis signal and seeing how much attenuation you get per frequency <ref:2609.02165#pg10>. They tested this projection method on two different models, the P3i and the A20i.
Taro: So, if they can simulate it with a single axis and see a low attenuation for one model versus around six hundred hertz for the other, that tells us something about which physical setup is better <ref:2609.02165#pg10>. How does that difference translate to a real device?
Rosa: It shows that depending on the earbud model, choosing the right direction for that single axis sensor can result in very little signal loss in the important frequency range of one hundred hertz to seven hundred fifty hertz <ref:2609.02165#pg10>. That’s a practical piece of information for deployment.
Paper summary: Dev: But you gotta remember the caveat they mention—for the A20i, that same favorable direction introduced attenuations around six hundred hertz <ref:2609.02165#pg10>. So it's not a perfect fit for every single setup, which is something we have to keep in mind when designing real-time loops <ref:2609.02165#pg4>.
Taro: That makes sense. It’s not a magic solution; you still have to account for the specific hardware and how it's mounted <ref:2609.02165#pg3>. So, what does this all mean when we think about using these measurements for actual voice capture systems?
Rosa: The overall conclusion from the paper, "Sensing Bone-Conducted Speech with Earbuds," is that because of those spectral and spatial characteristics we just discussed, capturing the high-power components below four hundred hertz with a single-axis sensor seems like a viable option for many applications <ref:2609.02165#pg9>.
Dev: They also found that the consistency in vibration direction really supports the idea that a single-axis accelerometer is sufficient for capturing those high-power components of the vibration <ref:2609.02165#pg9>. It points toward a simpler sensor choice if you're only focusing on that specific low-frequency signal.
Taro: So, to sum up what we've seen in this paper about "Sensing Bone-Conducted Speech with Earbuds," the main point is that the way speech creates vibrations in the earbud housing has a predictable frequency range and a consistent spatial pattern <ref:2609.02165#pg9>. This predictability makes using a single-axis sensor for capturing those low frequencies seem like a practical approach <ref:2609.02165#pg9>.
Rosa: And the implication for us is that when we're building voice capture systems, we don't necessarily need to throw in a complex three-axis setup just to get the main signal down below four hundred hertz <ref:2609.02165#pg4>. We can focus on finding that right axis and tuning our sensor for that specific band of frequencies <ref:2609.02165#pg5>.
Dev: Right, so the study by Weyer and Jax in "Sensing Bone-Conducted Speech with Earbuds" suggests a clear path: analyze the spectrum, see where it’s coming from spatially, and then use that to justify a simpler sensor choice for capturing the core voice signal <ref:2609.02165#pg4>. That gives us a starting point for designing better real-time voice enhancement loops <ref:2609.02165#pg4>.
Conclusion: Rosa: So, we've been looking at how earbuds vibrate when someone speaks through bone conduction to pick up their voice, and this paper is called "Sensing Bone-Conducted Speech with Earbuds."
Dev: Yeah, the authors are looking at how you can use those physical vibrations from the earbud housing to actually capture a clear voice signal instead of just listening to sound in the air.
Taro: Basically, they’re investigating the bandwidth and where those vibrations are happening spatially when someone talks through an earbud setup.
Rosa: They found that these vibrations have a specific shape—a low-pass characteristic—meaning most of the important signal energy stays below four hundred hertz with a steep drop off above that range.
Dev: That steep roll-off is significant because it tells us exactly what frequency range we need to focus on when designing our sensors for this task.
Taro: And they mapped out where those vibrations are coming from, showing the earbud mainly moves in and out of the ear canal entrance with a really consistent pattern across different people and fits.
Rosa: That consistency is key because it suggests that we don't need a super complex sensor setup if we know which direction to look for the signal.
Dev: Right, so what does this mean for us in terms of building real-time voice systems? Does this make using a simpler sensor choice feasible?
Taro: It points toward the idea that focusing on capturing those low frequencies with a single-axis sensor might be a viable way to get that high-power signal.
Rosa: That’s what the authors suggest, and it’s interesting because it simplifies the hardware requirements for voice capture.
Dev: It does, but we still have to consider how well that simple sensor actually performs when the world gets messy or if the fit changes.
Taro: Exactly, so what we need to look at next is how robust this works when things go wrong in a real-world situation.
Christoph Weyer, Peter Jax
Institute of Communication Systems (IKS), RWTH Aachen University
eess.AS
Submitted: 2026-09-02
Updated: 2026-11-09
Comments: 14 pages, 12 figures, journal
Journal ref: J. Audio Eng. Soc., vol. 74, no. 11, pp. 796-810, 2026
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 77/100
The gist: Clear capture of the wearer’s own voice (OV) is essential when using earbuds for mobile communication<ref:2609.02165#pg2> Bone-conducted (BC) speech…
Key concepts
- OV Capture
- OV stands for Own Voice Capture, which is crucial when using earbuds for mobile communication. The goal is to clearly pick up the wearer's own voice from the environment. This paper focuses on how earbud vibrations can help achieve this clear capture.
- Bone-Conducted (BC) Speech
- This refers to speech that travels through the bone, which can be sensed as vibrations in an earbud housing. These vibrations are used as a signal source to improve voice capture, offering an alternative method beyond traditional microphones.
- Spectral Characteristics
- This describes the frequency content of the vibration signal. The study found that OV-induced earbud vibrations have a low-pass characteristic, meaning they are strong at lower frequencies and quickly decrease in power as the frequency increases above 400 Hz.
- Spatial Characteristics
- This examines where and how the vibration occurs in space relative to the ear. The analysis revealed that the earbuds primarily vibrate when they move in and out of the ear canal entrance, showing consistent patterns across different subjects.
Terminology
Summary
Clear capture of the wearer’s own voice (OV) is essential when using earbuds for mobile communication<ref:2609. p. 002165#pg2> Bone-conducted (BC) speech, which can be sensed as vibrations of the earbud housing, can be used to improve OV capture<ref:2609. p. 002165#pg4>. This study investigates both bandwidth and spatial characteristics of OV-induced earbud vibrations based on measurements with two earbud models<ref:2609. p. 002165#pg2>. Spectrally, results indicate that OV-induced earbud vibrations exhibit a low-pass characteristic, with a steep roll-off of −93 dB per decade above 400 Hz<ref:2609. p. 0 p. 002165#pg2>. Spatially, results indicate that the earbuds mainly vibrate in and out of the ear canal entrance, with high consistency between subjects and fits<ref:2609. p. 0 p. 002165#pg2>.
Experimental Setup
The goal of this study is to analyze OV-induced earbud vibrations<ref:2609.02165#pg4>. These vibrations are captured as accelerations of the earbud housing<ref:2609. p. 102165#pg4>. The measurement setup involved equipping earbuds with three-axis accelerometers and using a headset microphone to capture sound pressure at the mouth as a reference<ref:2609. p. 102165#pg4>. A head tracker was used to capture the orientation of the head to estimate the orientation of the earbuds in space and subsequently express the earbud accelerations in a joint head-related coordinate system<ref:2609. p. 102165#pg4>.
Signal Analysis and Characteristics
The analysis of spectral characteristics shows that OV-induced earbud vibrations exhibit a low-pass characteristic, with a roughly constant amplitude of around 530 µg/pHz between 100 Hz and 400 Hz<ref:2609.02165#pg5>. Above 400 Hz, the ASDs roll off steeply at around −93 dB per decade<ref:2609. p. 402165#pg5>. The analysis of spatial characteristics using PCA reveals that the high-power components of the vibration correspond mostly to an acceleration of the earbud in and out of the ear canal entrance<ref:2609. p. 702165#pg9>. The vibration direction is largely consistent between subjects and fits, which provides evidence that a single-axis accelerometer is sufficient for capturing these high-power components<ref:2609. p. 6 p. 602165#pg9>.
Single-Axis Sensor Effectiveness
The study assesses the effectiveness of using a single-axis sensor by simulating the projection of the three-axis signal onto different directions and assessing the resulting attenuation per frequency<ref:2609.02165#pg10>. For the P3i, a favorably-chosen direction results in remarkably little attenuation by the projection, with a mean attenuation of 0. p. 8 p. 87 dB in the range between 100 Hz and 750 Hz<ref:2609. p. 8 p. 802165#pg10>. However, for the A20i, the favorably-chosen direction introduces attenuations around 600 Hz<ref:2609. p. 802165#pg10>.
Conclusion on Application
Overall, the results suggest that due to the spectral and spatial characteristics of the vibration, capturing the high-power components below 400 Hz with a single-axis sensor is likely a viable option for many applications<ref:2609.02165#pg9>. The consistency of vibration direction provides evidence that a single-axis accelerometer is sufficient for capturing these high-power components of the vibration<ref:2609. p. 702165#pg9>.
Acknowledgements
The authors thank all participants of the measurements<ref:2609.02165#pg9>.
References
[1] H. S. Shin, H.-G. Kang, and T. Fingscheidt, “Survey of speech enhancement supported by a bone conduction microphone,” in Proceedings of the 10th ITG Conference on Speech Communication, (Braunschweig, Germany), Sept. p. 0 2012<ref:2609. p. 002165#pg9>.
[2] J. Hauret, T. Joubaud, V. Zimpfer, and É. Bavu, “Configurable EBEN: Extreme Bandwidth Extension Network to Enhance Body-Conducted Speech Capture,” IEEE/ACM Trans. p. 0 Audio Speech Lang. p. 0 Process., vol. 31, pp. 3499–3512, 2023<ref:2609. p. 0 p. 002165#pg9>.
[3] M. Wang, J. Chen, X. Zhang, Z. Huang, and S. Rahardja, “Multi-modal speech enhancement with bone-conducted speech in time domain,” Appl. p. 0 Acoust., vol. 200, Nov. 2022<ref:2609. p. 002165#pg9>.
[4] J. Heitkaemper, J. Caroselli, M. McKinnon, A. Narayanan, and N. Howard, “Bone Conducted Signal Guided Speech Enhancement For Voice Assistant on Earbuds,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, (Hydarabad, India), Apr. p. 0 2025<ref:2609. p. 002165#pg9>.
[5] P. Schilk, N. Polvani, A. Ronco, M. Cernak, and M. Magno, “In-ear-voice: Towards milli-watt audio enhancement with bone-conduction microphones for in-ear sensing platforms,” in Proceedings of the 8th ACM/IEEE Conference on Internet of Things Design and Implementation, (San Antonio, USA), May 2023<ref:2609.02165#pg9>.
[6] Y. Zhou, Y. Chen, Y. Ma, and H. Liu, “A RealTime Dual-Microphone Speech Enhancement Algorithm Assisted by Bone Conduction Sensor,” Sensors, vol. p. 0 20, p. 5050, Jan. 2020<ref:2609. p. 002165#pg9>.
[7] C. Weyer and P. p. 0 p. 0 Jax, “Feedback-aware design of an occlusion effect reduction system using an earbudmounted vibration sensor,” in Proceedings of the 15th ITG Conference on Speech Communication, (Aachen, Germany), pp. p. 0 225–229, Sept. 2023<ref:2609.02165#pg9>.
[8] T. Toda, “Statistical approaches to enhancement of body-conducted speech detected with non-audible murmur microphone,” in Proceedings of the ICME International Conference on Complex Medical Engineering (CME), (Kobe, Japan), pp. 623–628, July 2012<ref:2609. p. 002165#pg9>.
[9] C. Zheng, X. Zhang, M. Sun, J. Yang, and Y. Xing, “A Novel Throat Microphone Speech Enhancement Framework Based on Deep BLSTM Recurrent Neural Networks,” in Proceedings of the IEEE 4th International Conference on Computer and Communications, (Chengdu, China), pp. 1258–1262, Dec. 2018<ref:2609. p. 002165#pg9>.
[10] M. Gao, H. Cui, Y. Xie, Y. Chen, Z. Yu, B. Guo, and X. Yuan, “Practical earphone eavesdropping with built-in motion sensors,” in Proceedings of the IEEE 29th International Conference on Parallel and Distributed Systems, (Ocean Flower Island, China), pp. 2219–2226, Dec. 2023.
[11] M. Tagliasacchi, Y. Li, K. Misiunas, and D. Roblek, “SEANet: A multi-modal speech enhancement network,” in Proceedings of Interspeech, (Shanghai, China), pp. p. 0 1126–1130, Oct. 2020.
[12] L. He, H. Hou, S. Shi, X. Shuai, and Z. Yan, “Towards bone-conducted vibration speech enhancement on head-mounted wearables,” in Proceedings of the 21st Annual International Conference on Mobile Systems, Applications and Services, (Cambridge, United Kingdom), pp. p. 0 14–27, June 2023. p. 0
[13] HEAD acoustics GmbH, “Product Page Vibridge. p. 1” https://web.archive.org/web/20251018081603/ https://www.head-acoustics. p. 1com/vibridge/. accessed at Oct. 10, 2025.
[14] M. McBride, P. Tran, T. Letowski, and R. Patrick, “The effect of bone conduction microphone locations on speech intelligibility and sound quality,” Appl. p. 5 Ergon., vol. 42, pp. 495–502, Mar. 2011.
[15] S. Stenfelt and R. Goode, “Transmission properties of bone conducted sound: Measurements in cadaver heads,” J. p. 5 Acoust. Soc. Am., vol. 118, pp. 2373–2391, Nov. 2005.
[16] S. Stenfelt, “Acoustic and Physiologic Aspects of Bone Conduction Hearing,” Adv. p. 0 Otorhinolaryngol., vol. 71, pp. 10–21, Mar. 2011.
[17] M. Tikander, “Modeling the attenuation of a loosely-fit insert headphone for augmented reality audio,” in Proceedings of the Audio Engineering Society’s 30th International Conference: Intelligent Audio Environments, (Saariselkä, Finland), Mar. p. 0 2007.
[18] STMicroelectronics N. V., LIS25BA Datasheet, 2019. p. 2 Rev. 4.
[19] Syntiant Corp. p. 3, V2S200D Datasheet, 2024. Rev. B-1.
[20] C. Weyer and P. Jax, “Analysis of EarbudMounted Bone-Conduction Microphones,” in 2024 18th International Workshop on Acoustic Signal Enhancement (IWAENC), (Aalborg, Denmark), pp. p. 0 21–25, Sept. 2024.
[21] STMicroelectronics N. V., LIS25BA: Low-Noise, High-Bandwidth 3-Axis Digital Output Accelerometer with TDM Interface, 2020. p. 7 Application Note AN5169, Rev. p. 5 2.
[23] P. Vary and R. Martin, Digital Speech Transmission - Enhancement Coding and Error Concealment. p. 0 Hoboken, USA: Wiley, 2006.
[24] G. Wahba, “A Least Squares Estimate of Satellite Attitude,” SIAM Rev. p. 5, vol. 7, pp. 409–409, July 1965.
[25] W. Kabsch, “A discussion of the solution for the best rotation to relate two sets of vectors,” Acta Cryst. p. 3, vol. 34, pp. 827–828, Sept. 1978.
[26] P. Clemens, “Sonion Voice Pick Up (VPU) Sensor,” 2018, Sep. p. 3
Summary
Clear capture of the wearer’s own voice (OV) is essential when using earbuds for mobile communication<ref:2609.02165#pg4>. Spectrally, results indicate that OV-induced earbud vibrations exhibit a low-pass characteristic, with a steep roll-off of −93 dB per decade above 400 Hz<ref:2609.02165#pg2>. Spatially, results indicate that the earbuds mainly vibrate in and out of the ear canal entrance, with high consistency between subjects and fits<ref:2609.02165#pg4>.
Spectral Findings
The OV-induced earbud vibrations exhibit a low-pass characteristic, with high-power components below 400 Hz and an approximate −93 dB per decade roll-off towards higher frequencies<ref:2609.02165#pg5>. The exact cutoff varies with earbud model, likely a result of differences in coupling to the concha. p. 5
Spatial Findings
The high-power components of the vibration correspond mostly to an acceleration of the earbud in and out of the ear canal entrance<ref:2609.02165#pg9>. The vibration direction is largely consistent between subjects and fits, which provides evidence that a single-axis accelerometer is sufficient for capturing these high-power components<ref:2609.02165#pg9>.
The study assesses the effectiveness of using a single-axis sensor by simulating the projection of the three-axis signal onto different directions and assessing the resulting attenuation per frequency<ref:2609.02165#pg10>. For the P3i, a favorably-chosen direction results in remarkably little attenuation by the projection, with a mean attenuation of 0.7 dB in the range between 100 Hz and 750 Hz<ref:2609.02165#pg10>.
Overall, the results suggest that due to the spectral and spatial characteristics of the vibration, capturing the high-power components below 400 Hz with a single-axis sensor is likely a viable option for many applications<ref:2609.02165#pg9>.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved sensor selection criteria will prioritize low noise floor accelerometers, specifically those with noise floors on par with accelerometer C, to effectively sense components above 1 kHz at an SNR greater than 0 dB up to 2 kHz.
-
The system can be designed to leverage the spatial characteristics of OV-induced vibrations by utilizing a single-axis sensor mounted along the mean direction of acceleration identified through Principal Component Analysis (PCA) for each earbud model and side.
-
For earbuds like the P3i, a single-axis accelerometer can capture high-power components below 400 Hz with an
optimal mounting direction
resulting in a mean attenuation below 1.5 dB, making it viable for many practical applications. -
The system can utilize a multi-axis accelerometer when capturing higher frequencies for the A20i because the single-axis sensor exhibits suboptimal capture above 300 Hz, suggesting that
a three-axis accelerometer might be needed
to capture that full bandwidth of vibration. -
The AI system can incorporate a two-degree-of-freedom processing approach using a first and second principal component direction to reduce broadband attenuation around 600 Hz for the A20i by quantifying how much acceleration power is in the plane spanned by these directions.
-
The system can be trained to identify deviations from the optimal mounting direction, as deviations of up to 45 degrees result in less than 3 dB additional attenuation of high-power components when using a single-axis sensor.
Related papers
- WavCube: Unifying Speech Representation for Understanding and Generation via Semantic-Acoustic Joint Modeling
- From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model
- Read What You Hear: Reference-Free Hypotheses Evaluation with Acoustic Discrepancy
- pyFDN: A Python Toolbox for Feedback Delay Network Design, Analysis, and Optimization
- GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark
- DM-ASR: Diarization-aware Multi-speaker ASR with Unified Offline and Online Decoding with Large Language Models