A quantitative evaluation of hearing aid algorithms is a crucial step for developers and clinicians. It allows a better understanding of the effect of different signal processing strategies in different conditions, i.e. by using different test signals and fitting options. The results can be used to prepare clinical tests and to select the best solution for our products. While this noble intention might be universally accepted, challenges arise when it comes to how to do the evaluation. This blog post explains the theory behind the output signal-to-noise ratio (SNR) measurements and proposes a method to evaluate Bernafon’s Dynamic Amplification Control™ (DAC™). The supplementary materials included with this blog can be used to reproduce our results and extend this technique to your own use cases.
Dynamic Amplification Control™
It is important to revisit the main principle of DAC™ before defining the test setup and interpreting the results. Bernafon introduced DAC™ with Zerena as a complement to the ChannelFree™ compression system. Compression typically defines amplification as a function of the estimated sound pressure level (SPL) without differentiating the type of the incoming signal, i.e. speech or noise. This missing information is provided by DAC™ which estimates the SNR. This SNR indicates if the signal that will be amplified consists of more speech or more noise. Speech will be amplified as programmed and can be verified with real ear measurements (REM), while gain applied to the noisy portions of the signal will be reduced with DAC™. In other words, amplification has a quantitative characteristic (based on the estimated SPL) and a qualitative characteristic (based on the type of the incoming signal).
The effect of DAC™ is measurable only if speech and noise are present in the test signal. It is also important that the speech signal contains natural pauses, i.e. locally low SNRs, to see how DAC™ adapts the amplification to fast and large variations of the SNR. The challenge, when we want to look at this effect, is to know how the hearing aid amplifies speech and noise presented simultaneously. This is possible with the inversion technique introduced by Hagerman & Olofsson (2004) that is used to measure the SNR at the output of the hearing aid.
Output hearing aid SNRs
The hearing aid output SNR provides valuable information when you prepare a clinical test. It allows us to define which listening condition could be used to detect a difference in technology with hearing aid users. It has been shown that an improved output SNR is associated with better speech intelligibility (Brons et al., 2014) and less listening effort (Sarampalis et al., 2009).
The inversion technique principle is to make two successive recordings with the selected signals you want to separate. While these signals are presented simultaneously, you can still adjust the level of each one to cover a defined range of input SNRs. For our test case, we will start with a speech and a noise signal. While the first recording (A) has both original signals, speech is inverted in the second one (B). In a post-processing phase, you can combine the recordings to isolate each signal of interest. The estimated level of the extracted speech and the extracted noise can be used to estimate the hearing aid output SNR.
Figure 1: Principle of the inversion technique to compute the SNR of two signals played simultaneously.
Experimental setup
We recorded the effect of DAC™ in our Bernafon Sound Studio with an artificial head. Amplification was set to match NAL NL2 targets for the N3 standard audiogram (Bisgaard et al., 2010). The Noah file with all fitting characteristics is provided with the supplementary materials.
Figure 2: Recording setup in our Bernafon Sound Studio with the KEMAR and both speakers used for the experiment.
The International Speech Test Signal (ISTS, Holube et al., 2010) was selected as our speech signal and played from the front speaker (at 0°). This signal has a number of advantages: it is easily accessible, it is commonly accepted as a standard test signal, it provides a steady noise that matches the long-term spectrum of the speech signal, and it presents natural speech characteristics with natural pauses, where we expect to see the effects of DAC™.
The noise is played from a speaker located at 30° and its output level is adjusted to cover a range from -5 up to +20 dB SNR with 5 dB steps. There might be some minor differences between the reference SNR measured in the free field and the unaided SNR due to the relative azimuth of both signals.
The estimated output SNR is a highly valuable outcome as it is computed over the entire test signal. It gives us a quantitative appreciation of what is happening in the long term with amplified speech and noise but doesn’t reflect qualitative local changes with different processing strategies. It might be interesting therefore to start the evaluation by visualizing the envelope of the extracted signals.
Visualizing how DAC™ affects the signal’s envelope
The envelopes of the extracted speech and noise are obtained with the inversion technique. We show here in Figure 3 a short utterance from the aided recordings at +15 dB SNR. The unaided signal, containing speech and noise, is plotted as a reference with the light gray dashed line.
Figure 3: Envelopes of the extracted speech and noise without (left) and with (right) DAC™. The speech signal is shown in the darker color and the noise in the lighter one. The unaided signal (with speech and noise) is shown in gray.
Extracted noise, represented with the lighter color, shows less amplitude variations over time when processed with DAC™ than with conventional level dependent compression. This might not be a big issue until speech dominates noise. But this effect is prominent during speech pauses, where conventional compression boosts the low-level signal, i.e. the low-level signal is mainly noise for a test at 15 dB SNR.
Naylor & Johannesson (2009) reported this effect for any compression architecture at positive SNRs. It can be explained by the fact that compression reduces the “distance” between the speech and the noise signal. This over-amplification during speech pauses might provide too much audibility of noise and might be perceived as being annoying. DAC™ was designed to mitigate this effect by adapting the amplification to the type of incoming signal. We can now quantify this effect by estimating the long-term output SNR.
Improved output SNR with DAC™
The output SNR is computed on the long-term signal as we make no assumptions about the importance of different parts of the test signals. The aided long-term output SNR is estimated for each tested condition defined by its unaided SNR and is presented in the Figure 4 below.
Figure 4: Results of the measurements without and with DAC™ with ChannelFree™ compression.
The difference between both curves depends on the unaided SNR, i.e. more benefit can be observed on the output SNR with DAC™ at positive unaided SNRs. With a positive test SNR of 19 dB (on the right), the improvement with DAC™ is about 2.3 dB SNR.
This improvement might not directly lead to a measurable benefit of intelligibility. Only noise is attenuated while speech audibility should be maintained. However, noise is much less annoying when it is stable and softer. Lesimple & Tantau (2017) showed that DAC™ improves the response time while preserving the speech intelligibility in a word recognition test. These findings are especially important as Houben et al. (2013) suggested that a faster response time during a speech test might reflect reduced listening effort.
Some supplementary materials before you get started
The supplementary material folder contains all the recordings and the Noah file if you want to reproduce this experiment. Scripts for extracting signals, computing output SNRs, and plotting the figures can be used and modified to suit your own needs. They are written in R which is an open source software available free of charge.
All supplementary materials are here for you in this zip folder.
Here are some additional tips if you want to perform your own measurements:
- Take some time to calibrate and adjust the level of the signals, especially if the devices are fitted with compression. Some systems might be designed with different compression thresholds that will create differences in the recordings that are difficult to interpret.
- Some feedback cancelling systems induce a change in the phase of the signal. This is a problem for the inversion technique as this shift can’t be controlled. It is therefore better to disable the feedback canceller and use a closed fitting to avoid any feedback problems. If it is still an issue, then decrease the gain or use some putty to seal around the instrument.
- Test variation with another noise type can be used for further scenarios. Thus, SNR estimation might be challenging when the noise and speech don’t have the same long-term spectrum and when the hearing aid has frequency specific amplification. Therefore, it is recommended to use test signals with the same frequency components.
This blog post can be used as a basic introduction to how to measure the effect of different algorithms like DAC™. It should support clinicians and researchers to reproduce our results and extend this technique to their own test cases, e.g. evaluating noise reduction or directional microphone systems. I believe that a better understanding of hearing aid technology can only help clinicians to improve fitting and fine tuning with the goal to increase the satisfaction of the hearing aid user.
References
Bisgaard, N., Vlaming, M. S., & Dahlquist, M. (2010). Standard audiograms for the IEC 60118-15 measurement procedure. Trends in Amplification, 14(2), 113-20.
Brons, I., Houben, R., & Dreschler, W. A. (2014). Effects of Noise Reduction on Speech Intelligibility, Perceived Listening Effort, and Personal Preference in Hearing-Impaired Listeners. Trends in Hearing, 18, 233121651455392.
Hagerman, B., & Olofsson, A. (2004). A method to measure the effect of noise reduction algorithms using simultaneous speech and noise. Acta Acustica United with Acustica, 90(2), 356–361.
Holube, I., Fredelake, S., Vlaming, M., & Kollmeier, B. (2010). Development and analysis of an International Speech Test Signal (ISTS). International Journal of Audiology, 49(12), 891–903.
Houben, R., van Doorn-Bierman, M., & Dreschler, W. A. (2013). Using response time to speech as a measure for listening effort. International Journal of Audiology, 52(11), 753–761.
Lesimple, C., & Tantau, J. (2017). Benefits of dynamic amplification control in complex listening environments. [White Paper]. Retrieved March 2019, from www.bernafon.com/professionals.
Sarampalis, A., Kalluri, S., Edwards, B., & Hafter, E. (2009). Objective Measures of Listening Effort: Effects of Background Noise and Noise Reduction. Journal of Speech, Language, and Hearing Research, 52(5), 1230–1240.