TL;DR
This paper investigates how well people understand speech delivered through bone-conduction headphones when there's competing noise, especially when the speech is coming from a specific direction. They found that bone-conduction headphones perform nearly as well as traditional headphones, suggesting they're a viable option for applications like virtual reality or navigation aids for the visually impaired.
What Problem Does This Paper Address?
Imagine you're using a navigation app that speaks directions, or you're immersed in a virtual reality game. For these applications to be effective, the sound needs to feel like it's coming from a specific direction, and you need to understand it clearly, even with background noise. Traditional headphones cover your ears, which can be inconvenient or even dangerous if you need to hear your surroundings. Bone-conduction headphones offer a solution by transmitting sound through vibrations to your skull, leaving your ears open.
However, there's a question about how well these headphones work, especially when the sound source is localized (meaning it seems to come from a specific direction) and there's competing noise. A key concern is "cross-talk" – vibrations from the bone-conduction transducer might travel to the opposite side of your skull, potentially interfering with the intended sound. This paper aims to quantify the speech intelligibility differences between bone-conduction and traditional headphones under these challenging conditions, specifically when the target speech is localized diagonally.
Methods
The researchers used the Japanese Diagnostic Rhyme Test (DRT), a standard method for measuring speech intelligibility. In this test, participants hear a word and must choose it from a pair of rhyming words that differ by only one sound. They used 120 words from the DRT and tested them with both standard (air-conduction) and bone-conduction headphones.
Crucially, they simulated virtual acoustic environments. The target speech was localized at a 45° azimuth (to the right of the listener), while competing babble noise was placed at various other locations around the listener. They tested three different signal-to-noise ratios (SNRs): 0 dB, -6 dB, and -12 dB. To ensure fair comparison, they carefully adjusted the perceived loudness of the headphones to match a reference loudspeaker.
Datasets
The paper utilized a babble noise sample from the Signal Processing Information Base (SPIB) at Rice University. This sample consisted of 100 simultaneous speakers. The target speech was recorded from a single female speaker. The Head-Related Transfer Functions (HRTFs) used for spatialization were obtained from the Knowles Electronics Mannequin for Acoustic Research (KEMAR) database.
Key Insights and Ideas
- Bone-conduction headphones are surprisingly robust: Despite concerns about cross-talk, bone-conduction headphones performed nearly as well as traditional headphones in most scenarios.
- Diagonal localization is key to cross-talk effects: The subtle intelligibility differences observed between headphone types were most pronounced when the target speech was localized diagonally (45°). This suggests that the physical path of vibrations plays a significant role.
- Nasality is particularly sensitive: The intelligibility of speech sounds involving nasality (like "m" or "n") was more affected by the competing noise and headphone type than other phonetic features. This points to specific acoustic properties that are more susceptible to interference.
- Cross-talk creates a phantom sound source: The authors hypothesize that cross-talk in bone-conduction headphones can create a perceived sound image in the opposite ear, especially when the target speech is not directly in front. This "phantom" source can interfere with the intended speech, particularly at lower SNRs.
Key Results
- Overall, the average correct response rates for bone-conduction and air-conduction headphones were very similar across different SNRs and noise locations.
- At the lowest SNR (-12 dB) and when the target speech was at 45°, bone-conduction headphones showed a statistically significant decrease in intelligibility compared to air-conduction headphones for certain noise azimuths (-45°, 0°, and 45°).
- This intelligibility degradation was most prominent for nasal sounds, with correct response rates dropping noticeably for bone-conduction headphones under the challenging -12 dB SNR condition when the target speech was at 45°.
- When the target speech was directly in front (0°), no significant difference was observed between headphone types, even at -12 dB SNR.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Limited azimuths tested: They only tested the target speech at 45° and 0°. Future work should explore other azimuths, particularly 90°, and investigate scenarios where the target speech is in the left hemisphere.
- Bone-conduction noise: In this study, both the target speech and the competing noise were delivered via bone conduction. Real-world applications will likely involve competing noise from loudspeakers (air conduction). Future experiments should incorporate this more realistic scenario.
- Single speaker: The study used only one female speaker for the target speech. While they cite previous work suggesting minimal speaker dependency for intelligibility, testing with multiple speakers, including males, would strengthen the findings.
- Investigating other attributes: The reason why nasality was uniquely affected needs further investigation. Exploring other phonetic attributes and their susceptibility to cross-talk is also warranted.
Summary updated