TOWARDS AUDITORY ATTENTION DECODING WITH NOISE-TAGGING: A PILOT STUDY

Introduction

People suffering from hearing loss often have great difficulty in scenarios in which multiple individuals are speaking simultaneously, known as the ‘cocktail party scenario’, something which normal hearing persons have no difficulties with . In these scenarios, hearing aids are not able to provide a good solution, as even though they are capable to suppress background noise, they are less capable of suppressing the unattended speakers.

Some hearing aids attempt to mitigate this problem by using a heuristic, for example by enhancing the loudest or closest speaker, or the one who stands right in front of the listener.

Unfortunately, these heuristics often lead to selecting the wrong speaker in real-life scenarios.

Research Question

This work aimed to investigate fundamental insights in the application of noise-codes for auditory attention decoding. To investigate the speed of the models, sliding decision windows of length \u03c4 ranging from 1 to 60 s were used during testing.

Future studies need to investigate, if our observations generalize also to parallel stimulus presentation protocols.

Methodology

Using such a hybrid approach, Geirnaert and colleagues achieved a remarkable performance using canonical correlation analysis (CCA) to decode the attended speaker. This poses a significant limitation for real-world scenarios where fast speaker detection is crucial. Framing the speaker decoding problem as detecting which of several stimuli a person is attending to, another paradigm from the brain-computer interfacing (BCI) field recently reached remarkable performances.

Study Design

A c-VEP is the EEG response to pseudorandom visual stimulation sequences where stimuli are watermarked using noise-codes, a protocol called noisetagging .

Such c-VEP BCIs have been reaching state-of-the-art performances up to 100 % clas-sification accuracy using 1-4 s decision windows or recently even within 300 ms and a high number of stimuli, 29 and 40, respectively.

Results & Findings

These hearing aids aim to identify the attended speaker from neural activity, and correspondingly enhance this speaker’s audio signal whilst simultaneously suppressing the other speakers and background noise. Another approach is forward modeling or encoding, in which the objective is to predict the neural response from the speech envelopes via an encoder, and to compare these against the EEG .

• These hearing aids aim to identify the attended speaker from neural activity, and correspondingly enhance this speaker’s audio signal whilst simultaneously suppressing the other speakers and.
• Another approach is forward modeling or encoding, in which the objective is to predict the neural response from the speech envelopes via an encoder, and to.
• Presenting audio from two simultaneous speakers, a mean accuracy of about 85 % was reached using 30 s decision windows.
• However, when decreasing the decision window length to about 10 s, the accuracy quickly dropped to below 80 %.
• Firstly, we aim to assess the feasibility of decoding the code-modulated auditory evoked potential (c-AEP), the response to auditory noise-tagging.

Practical Applications

Furthermore, this exploration may pioneer a novel research avenue for the application of code-modulated responses in the auditory domain, a domain that has seen limited application compared to the visual modality, as so far only one study attempted this . It could be speculated that this is due to the modulation adding distinctive uncorrelated high-frequency content.

Future work could assess the perception thresholds for modulation depths, to obtain the least intrusive protocols for high usability.

Firstly, eCCA emphasizes global and higherorder activity associated with the speech envelope, while rCCA may focus more on early sensory responses evoked by the noise-codes.

Figures Explained

The paper’s visual material highlights the workflow and the main system components.

• Figure 1 :: Figure1: Visualization of three different modulation depths using noise-tagging. Depicted is the unmodulated audio, i.e., 0 percent (blue), and 50 (gold) and 100 (brown) percent modulated audio. Additionally, shown are the smoothened noise-tags used for modulation (black). Audio was amplitude-modulated by multiplying with the noise-code, retaining full audio amplitude when the code is 1, while only a percentage when it is zero. Therefore, the noise-code for 50 percent modulation ranges between 0.5-1, instead of 0-1 for 100 percent modulation. To ease comparison, we added the original audio (light gray) at the back of the modulated audio.
• Figure 2 :: Figure 2: Decoding accuracy across decision window length and modulation depth. Depicted is the grand average classification accuracy across decision window length \u03c4. Colored lines represent the five modulation conditions: 100 (blue), 90 (orange), 70 (green), 50 (red), and 0 (pink). Solid lines show the performance of rCCA and dashed lines for eCCA. The dashed horizontal gray line depicts theoretical chance level accuracy (50%).

Conclusion

Therefore, the noise-code for 50 percent modulation ranges between 0.5-1, instead of 0-1 for 100 percent modulation. The 100 modulation performed worse than 70 modulation but overall better than 50 modulation.

For modulation conditions 70 and 50, rCCA reached overall on par performances as eCCA.

Therefore, we used a 20 Hz lowpass and a higher sampling frequency.