I’m curious about using PEQ to compensate for changes in my hearing attributed to sound induced hearing loss. Apparently there are consequences to decades of loud music, scuba diving, and riding motorcycles. Recurring ear infections and temporary surgically inserted tubes probably didn’t help either. Ah, well.
So I have curves from Audiometry tests and I know pretty well where things are. This picture below isn’t my data but it’s a decent know representation to work with.
I’m exploring using PEQ to adjust for this specifically for headphone use. I also use convolution filters so that’s a factor to consider.
Is something like this the way to go?
Create a Procedural EQ filter.
Add a Parametric Equalizer operation which applies to “left” only.
Create a set of peak/dip filters that adjust from the measured points to the yellow line. For example, a 500Hz that reduces by 20dB and has a Q of 1/2 the distance to the next filter’s target.
Fill in the missing points on the upper and lower end by extrapolating.
Create a similar operation for the right side.
Is the best route? Peak/Dip with Qs as described. Is there a different approach?
Am I correct in assuming this can supplement convolution and, if so, does it matter if it goes before or after? I assume it has to go after.
Answers to any questions I haven’t asked would be very appreciated. Thanks!
Just curious, what are you using the convolution filters for? It would be possible to adjust those filters to incorporate the PEQ you are looking for and reduce the amount of processing. If you want to keep them separate, the order shouldn’t matter, provided that they don’t cause clipping.
Speaking of clipping: ideally, the yellow line should be chosen so that there will be little or no boosting and mostly attenuating. In this particular case, the difference between highest and lowest levels is over 60dB. That’s a lot to compensate for and would either greatly reduce the volume or risk clipping, so I guess you need to experiment a little and find a good compromise.
Convolution to - at least in theory - adjust to Harman but my recent, and first, trip to the audiologist shed some light on what’s going on with my hearing so maybe time to rethink.
I assumed that convolution to Harman + a custom PEQ to adjust for my personal hearing loss would get me to an approximation of what a person without loss would hear with a convolved bitstream.
In practice, I like the idea of keeping convolution separate from personalized PEQ. It’s easier to implement, the PEQ work can be re-used across headphones, and I can toggle convolution on/off without impacting personalization.
I’d have been tempted to bundle them in some combined convolution, but your reasons for not doing so are pretty convincing. Filters are two channel so varying the L/R convolution is possible. PEQ is easier to play with and test.
Some information
This is the result of tonal audiometry which refers to and tests the sound in the human speech band!
To tell us anything about your hearing, you need to have tests done:
Thanks, @Maxxim_M. I’ve done that. I have my results in hand. I posted an image of AC data from the internet instead of my own, but I have my own and they have some characteristics in common with what I posted.
I understand that the results are in db HL. Why would raising or lowering at specific frequencies be ineffective? Is there a short answer to that question or is a complex psychoacoustics issue?
If someone with AC results such as the results in the image I posted were to increase the decibel level of the upper mid range, would that help in any way? Would it help to increase the left more than the right to compensate for the difference?
I’ve been playing with PEQ - it’s not an easy journey. I was considering giving convolution filters a shot but the adjustments are large and it’s hard to imagine being able to do any of this without clipping.
It would help to understand your comments more. Thanks!
@Marian, @killdozer - can you point me in the right direction generating convolution filters? Can something like Deconvolver work or am I better off with Python? I don’t do much Python but if that’s the route to go, I bet ChatGPT and I could do it in an afternoon.
I have mild, recurring tinnitus, too. It’s annoying but, at least in my case, not as crippling as it is for some folks.
I went to an audiologist to get custom earplugs made and to get my hearing tested. One pair of earplugs for motorsports that fit comfortably under a helmet. One pair for concerts with swappable filters. I’m doing a multi-week motorcycle trip in Norway this summer and I have a concerts coming up (Nick Cave, if I can can get tickets on Monday, Chemical Brothers with tix already in hand, more). I don’t want to give that stuff up but I also don’t want things to get unnecessarily worse. I’m already pretty clearly on a path to hearing aids.
And now I have more context for what you’ve meant when you’ve commented that you’ve pre-convolved all or part of your library. This is definitely enough to get me started.
Glad this got you started. If it’s not too much to ask, please report back when/if you find a solution. You probably need to reduce the amount of correction in some frequencies to avoid clipping, but keep in mind that an occasional clipping is not the end of the world and will probably go unnoticed - unless you’re constantly watching the clipping indicator. High frequencies are usually at a low enough level so they could be boosted more.
Regarding db HL, it is basically db SPL that is weighted using the equal loudness curves, so that they become flat lines. That should not change the relative amount of hearing loss too much, so I think those diagrams are good guides.
Hello.
I will try to answer, Yes it is a complex issue. We have to remember that in audiometry we measure only a part/section of the sound band, but what about the rest ?
How does your hearing behave below and above the measured bandwidth? Note that only a few of the frequencies are tested, what about the rest? Intermediate frequencies and harmonics? An EQ is not perfect, so the result of its work will raise or lower the sound level not only at the indicated frequency, but also at the secondary frequencies and thus affect the harmonics. We often focus on raising our subjective evaluation of the quality of the equipment by referring to its price and our subjective sound perception preferences. No one thinks about research (What they hear and how they hear it) We talk about Natural Homogenous sound etc. Not realising that even with healthy hearing the reference range of the norm is 25dB HL, a lot, not much? Think of a situation where : you have a result 05/0 2.5/20 8/0 and 05/25 2.5/0 8/25. ( This is a shortened description, draw this graph on the picture you attached ) both are within the HL norm (i.e. correct peripheral hearing) and you will hear something different for each graph. If you have questions ask. I don’t want to lecture here, so I’m trying to simplify.
I appreciate your response. You point out that there are objective issues (e.g., small set of sampled frequencies, EQ induced distortion, etc.) and subjective issues (e.g., if I could “normalize” to a specified curve, would it sound “good’” or “better”). I understand these categories of issues and it’s helpful to hear you articulate them. Like, @killdozer, I’d buy a front-row ticket to a lecture on this general topic.
I’m curious if you think there is a path to something that people with small to moderate hearing loss might benefit from or enjoy. Consider a hypothetical tool that could take audiometry data as input, and produce a convolution filter as output. In principal, this seems similar to what AutoEQ does - it’s essentially an inversion of a curve which is subsequently manipulated to a target.
Is there reason to believe that, generally speaking, a tool like this might produce positive results for headphone users?
There are legitimate obstacles. Perhaps just these three:
Frequency range. Headphone samples cover 20Hz to 20kHz, audiometry tests appear to typically cover 250Hz to 8kHz.
Sample density. Perhaps effectively 696 samples for headphones? For audiometry, it’s on the order of 9.
Variance from target. “Moderate” hearing loss captures a large range - adjustment constraints (clipping/distortion) are likely to be more of an issue when adjusting for hearing loss than when adjusting for headphone variance.
Are these the categories? Are these just areas in which someone trying to work on a solution here would have to be clever (issues 1 + 2) and make compromises (issue 3)? I’ll stretch a bit here and say that you actually implied with your post that the goal doesn’t need to be perfection, it just has to be better in some ways, and hopefully not worse in any ways, than what someone with a bit of hearing loss is experiencing without it.
I’m curious what you and others think. I’m glad @Marian and @killdozer are on this thread - those folks know far more about audio science than I do.
What you can do for ranges that are not covered is simply extend the last values. For example if it’s 0dB for 500Hz and 20dB for 8kHz, use 0dB for 0-500Hz and 20dB for 8-20kHz. Or gently let the response go back to unity in the uncovered ranges.
I don’t think this is an issue. As you said, the goal is to make it better, not perfect.
It depends on the headphones. Also, when you combine the two, it may make things better or worse. You can clamp overall corrections to a certain maximum, say +20dB. It would not make the result “flat”, but it should improve it.
That’s what I was thinking. When I have the time, I might start by exploring what ChatGPT can do.
I pasted my audiometry results into ChatGPT 3.5. Literally just the tabular data just two rows of tabular data - the frequency column headers, and a row called “AC” with the results. I wrote, “This is audiometry data from a hearing test. Can you interpret it?” and got a fairly complete explanation and analysis.
I then wrote “I am using Roon as a music playback system. Roon allows me to use a parametric equalization filter to modify sound. Can you tell me the values to use to set up a parametric equalizer which adjusts for the audiometry data I provided?”
And it did. Including Roon specific instructions about how to about creating the filter, what values to use, what Q to start with, concerns about clipping/distortion. That approach ran out of gas but it would interesting to see what ChatGPT 4 can do with a small data set and a prompt requesting an expansion of the set which extends the margins and interpolates smooth curves. Hard to guess what it will do. If that doesn’t play out, I’ll do it myself or have one of my kids do it
Exactly. It’s worth looking at this to help with both reduced response and to normalize right and left.
Hello.
Thank you all for your interest in the topic, as this is one of the key areas of my research. Obviously, few people are interested in ‘knowing themselves’ and choose material values, such as buying a new cable. A new purchase is physical, hearing tests are not. By doing the research you get nothing physical except awareness. If I may, I will add, to illustrate the problem that -10 dB HL versus 20 dB HL is a difference that can be expressed as follows. Hearing, -10 dB HL is 1000 times more sensitive than the hearing of a person with 20 dB HL. That is to say, ‘‘The norm’’ is tracksuit trousers with elastic at the waist almost everyone can fit into, ‘‘Audio adaptation’’ is a bespoke tailor-made suit. You ask if it is possible to make a tool based on audiometric data? I will answer it is possible. Will it affect the quality of how you perceive sound? Of course it will! Even a person with normal hearing ( we are talking about the norm ) will hear a spectacular difference without needing hearing damage.
The adaptation itself, as I indicated above, is between 1 and 1000 in the normal area!
And the effect itself will be better when you input more data. You can extend audiometry not only to speech frequencies, but to the full range of the audible sound spectrum which will give you a range with a margin of 10Hz - 100 000 Hz yes 100kHz. You can calibrate not only audiometric headphones but also Hi-Fi headphones (I use Meze Empyrean myself). The physical aspect of the ear is only one piece of the puzzle, because even if you hear, the brain has to process it for you to understand what you hear. There is a fundamental difference between listening and hearing! Getting back on topic, you don’t need to create special boxes. It is enough to have a well-researched and imaged organ of your own hearing from the ear to the auditory cortex in the brain. I would add that this can be tested and the test is cheaper than the cable I mentioned The next step is to determine your hearing comfort level on the scale of the norm, and then make adjustments to your audio. As I wrote in the first post, this takes time, research and work.
( Of course I can write on this topic, but as you can see it is broad so it will be better to focus on individual issues, ask, to the best of my knowledge and experience I will be happy to answer )
Hello.
I would like to ask that we talk for a moment about @Marian’s solution .
‘’ What you can do for ranges that are not distributed is simply make the next values available. For example, if it’s 0dB for 500Hz and 20dB for 8kHz, then 0dB for 0-500Hz and 20dB for 8-20kHz. Or the response delay returns to unity in the original ranges.‘’
This solution is ingenious in its simplicity and could work if our hearing was linear and we could assume that everything below and above the damaged range is also damaged.
A little knowledge about the non-linearity of the ear: (scientific evidence few but spectacular)
doubling the strength of the wave acting on the basilar membrane of the organ of Corti does not produce a doubling of the discharge in the auditory nerve (less than a doubling of the discharge)
Adding a second tone that differs in frequency lowers the response to the first tone - 3. Playing two tones, e.g. 1000 and 1200 Hz, causes a third tone (in this case 800 Hz) - the so-called Cubic Difference Tone.
The brain adds (restores) the missing fundamental frequency.
Probably similarly, the “quintina” (“GHOST SOPRANO” ) of the fifth female voice) in Sardinian vocal music is created as an illusion!
So, selective hearing loss manifests itself as a lack of perception of specific frequencies. Specific and not all of them. Going further, the lack of perception of specific frequencies implies an additional disturbance in the perception of music perception. Further, there is the phenomenon of stimulation of hair cells that are in close proximity to naturally stimulated cells. Adding 20dB aggravates hearing damage over a wider range of frequencies ( researched ).
Examples could be multiplied, but let’s stop at these few.
As you can see this solution will not quite work.
The next step is to determine your hearing comfort level on the scale of the norm, and then make adjustments to your audio. As I wrote in the first post, this takes time, research and work.