distortion of interaural phase/time difference in openMHA output

[Sathish_kumar]

Dear forum members,
I am writing this post to seek your valuable opinions regarding the distortion of Interaural Phase Difference (IPD) or Interaural Time Difference (ITD) in the output of hearing aid (HA).

Numerous research studies have highlighted the distortion of available ITD or IPD information in HA output through objective verification methods. However, justifying the reasons for this distortion remains challenging due to the lack of technical details on the signal processing involved in commercial HAs. I assume that the distortion of IPD may be attributed to the nature of the filter bank used to implement multi-channel compression.

For instance, the Resound hearing aids utilize a frequency warping filter bank, incorporating non-linear frequency weighting to mimic human cochlear processing. Conversely, Infinite Impulse Response (IIR) filter banks have been employed in HAs due to their computational efficiency. Please correct me if I am mistaken, I think both frequency warping and IIR filters exhibit non-uniform phase delays depending on frequencies.

Recently, I ran the openMHA code available at: https://github.com/HoerTech-gGmbH/openM ... ession.cfg. According to the script, openMHA processing employs a Finite Impulse Response (FIR) filter bank with a linear phase delay. Consequently, it should preserve the IPD. However, the IPD in the openMHA processed output appears distorted. Unfortunately, due to limited information on the FIR filter bank used in openMHA, I'm unable to figure out the possible reasons for IPD distortion.

I kindly request forum members to share their thoughts and comments on the potential reasons for IPD distortion.

[nmichael]

Satish,

I ran checked this example and didn't see any phase difference between the channels. I checked this in an informal but thoroughly convincing way (to me at least). I made an input wav file that had the same signal in both channels then looked at the output wav file and saw that the channels still matched each other. The input wave file supplied with the example has different signals in each channel.

It's my understanding that the filterbank here is formed directly from the FFT of the signal so it should have a linear phase response. Even if the phase response was nonlinear I don't think this would produce interaural phase mismatch would it? As long as both channels are processed with the same algorithm the outputs should be identical for identical inputs. On the other hand as you correctly point out, the ReSound type frequency warping can cause phase distortion.

Nick

[Sathish_kumar]

Dear Michael,
Apologies for the delayed response. I agree with your observation that the openMHA algorithm preserves the ITD information when there is no difference across the channels (i.e., ITD = 0). However, I have noticed that when the ITD is introduced in certain instances, it tends to cause distortion.

I ran a simulation using the following method.
1. Introduction of ITD: I spatialized the audio signal using the Oldenberg HRTF database, accessible at https://uol.de/mediphysik/downloads/hearingdevicehrtfs. In this process, ITD is minimal for stimulus arriving from the front (i.e., 0 degrees), and gradually increases, reaching its maximum when the stimulus is from the side of the head (i.e., 90 degrees). The spatialized stimulus also introduces intensity difference across the audio channels, called Interaural Level Difference (ILD).

2. ITD calculation: I used the "xcorr" function in MATLAB to calculate ITD, This function utilizes the cross-correlation method to measure the similarity of two series as a function of the displacement of one relative to another.

Please refer to the attached figure, it shows that the openMHA processing results in changes to the available ITD between 30 and 80-degree azimuths.

simulation.png (44.75 KiB) Viewed 22571 times

[hendrikkayser]

Hi,
I assume that the differnces in the ITDs for angles larger than 30° are caused by the amplification and its effect on the results of xcorr. The gaintable used in that example is not symmetric for left and right channels and it has a compressive characteristic. This yields direction-dependent differences in the left/right spectra if ILDs get larger. This effectively changes the frequency weigthing in the computation of the broadband correlation function and potentially leads to the observed results.
I would expect to see no differnces for a constant gaintable.

[Sathish_kumar]

Yes, there are no changes in ITD when the constant gain table is used.