Greetings! I am Sharur, the founder and CEO of PEQdB! I am studying and researching audio science at Stanford's Center for Computer Research in Music and Acoustics. I saw a forum member named Lieqlien mention PEQdB in another thread, and I thought it would be helpful for professional audio engineers to know about our free tool. We engineered it to maximize sound quality through my accumulated psychoacoustics knowledge and utilizing cutting-edge data science and machine learning principles.

For the speaker calibration software, most speaker/room calibration software such as Sonarworks SoundID, Genelec GLM, and Dirac Live measure the left and right channels separately to calibrate them individually to a target curve. However, sound sources couple to room resonances differently when operating in unison versus in isolation, meaning that the measured response of the speakers playing at once will be considerably different than when playing them separately, particularly in the lower frequencies, which is why PEQdB measures both channels playing at once for its calibration. Users should use a microphone with an omnidirectional polar pattern below at least 500 Hz with a flat response for optimal results. Instead of forcing a user to take measurements at multiple different locations to form a spatial average, we let users do it in one go by playing periodic pink noise out of both channels and moving the microphone around the listening position to capture a spatial average in a few seconds via the moving microphone method. PEQdB only calibrates below 500 Hz while referencing the in-room measurement taken by the user because it is a common figure for the actual "room transition frequency" in typical listening rooms. Still, users can set this cutoff frequency between 250-500 Hz under the advanced options. This frequency defines the point at which the sound coming directly from the speaker becomes perceptually dominant over the combined sound energy, which is primarily influenced by room reflections. Above the set transition frequency, if a user selects their speaker model, we calibrate the on-axis response to be perfectly flat based on anechoic measurements known as Spinoramas. Finally, the user performs a listening test where they rate various low-frequency equalizations based on their subjective sound quality for 15 trials, after which our machine learning algorithm generates the most scientifically optimized equalization files possible.

You can perform a listening test with over 7000 headphones for headphone software and receive a customized equalization profile personalized to a user's taste. The "Ranking" mode was our first method and involves nine variable parameters. There are three filters: low-shelf filter, ear-gain peaking filter, and high-shelf filter. Each filter has variable frequency, Q-factor, and gain. Listeners rate each sound sample on a scale of -5 to 5 based on their perceived sound quality. Forty trials were required to minimize the error rate acceptably. The A/B Slider is our latest method and involves a neural network trained on data from approximately 1000 completed "Ranking" mode tests. For each trial, A and B are two different equalizations, and listeners can linearly interpolate between them on the fly to generate virtually infinite equalization profiles. With this method, a user can receive a perfect equalization profile in as little as 1-3 trials.

Our graphing tool shows the graphs of over 7000 earphones and the PEQdB Over-Ear and In-Ear target curves, the average preference curves derived from thousands of completed listening tests through our software. These curves far exceed the performance of target curves from companies like Harman, which were created with inferior methodologies, such as using only two variable parameters versus our nine. While Harman's research used inconsistent methodology and showed significant differences between in-ear and over-ear target curves, our data shows the differences are very subtle.

Lastly, we have a Left/Right ear balancer where listeners listen to ten band-passed ascending in frequency ranges and drag a slider for each to generate an L/R parametric equalization file. This helps address imaging issues for listeners whose ear balance is inconsistent across the frequency range.

Try our solutions on https://peqdb.com

 

Last edited: Mar 30, 2025

As a bonus, if you are curious about the audibility/preference for absolute polarity, you can perform an absolute polarity preference test on our website. From the thousands of trials completed, the confidence % that correct absolute polarity is preferred over inverted absolute polarity with asymmetric low frequency stimuli is 100%

 

Can you just put good monitors in your studio and calibrate to your own taste? It isn't really that difficult.

Not to mention to calibrate an headphone, how silly it is.

Good luck anyway with your biz

 

What do you consider good monitors and what is your calibration methodology?

The responses of headphones are all over the place with virtually none offering truly neutral sound out of the box.

 

Only top end expensive headphones and monitors available on the app to choose from. So this is not for "poor" people, right?

 

My calibration methodology is very easy, I put a chair in front of the monitor and if I like how they sound that's all.

 

Last edited: Mar 30, 2025

The list of equipment is dependent on what has been publicly measured with industry standard test equipment. For speakers, the Klippel Near Field Scanner costs about $100,000, so this makes more sense, but there are plenty of inexpensive headphones available for selection like the Tangzu Wan'er. Regardless, you do not need to select a speaker to perform the speaker calibration.

 

I expect you'd be highly impressed after doing the speaker calibration and hearing the improvement below 500 Hz

 

Consider that any improvement, most of the times, can be misleading too. I would suggest to invest more effort in learning how a monitor/headphone sound then calibrate it (put an EQ curve at the very final stage).

 

this is so cool of you!

 

While there is some validity to this approach, it is more likely to result in poor mixes versus using a calibrated system since you have to wire your brain to adapt to the colorations of your specific system, meaning you have to relearn everything if you want to mix on a different system. It's best to mix with a professional grade system so that it translates perfectly to other professional systems.

 

Hi Sharur, it's great that you're putting all those tools out for free. There's a number of issues with your approach that I'll try to detail, and I'll try to suggest on ways to improve on them.

About the headphones calibration suite :

1. The source of the measurements. The data is taken from multiple actors from the community (with their permission I hope ??), but with different measurement rigs. Different IEC711 clones will have different responses. This introduces some irregularities in the measurements, especially above 10kHz.
2. The whole calibration process is quite messy. The default "reference" tracks aren't really reference quality, and letting the user upload their own tracks will lead to inconsistent results, as it will completely depend on the tracks they choose.
3. Nothing above 10kHz is compensated for, which should not be ignored.

So here's what you can do :

1. Build your own measurement database, or send a reference unit to every provider of your measurements so that you truly know what the response of their rig is.
2. Don't use blind tests with A/B testing or ranking. Your PEQdb target is already quite good, close to the ideal (tilted avg DF curve with bass boost), so the adjustments should be very simple. Level and frequency of the bass boost, and overall tilt (it's actually only what Sonarworks offered before they introduced custom targets). I recommend that you use a combination of true reference tracks (neutral orchestral recordings, speech and foley material, solo acoustic instruments...) so that you bypass the artistic intent of the engineers, and then some selected music tracks that are highly regarded for their engineering (Dire Straits, Random Access Memories, Thriller, RATM's first album, Billie Eilish, Al Schmitt and Serban Ghenea mixes... to name a few), and let the user upload a few tracks. Then, give the user control over the three parameters I talked about. It doesn't need to be a blind test, just don't display the values when adjusting so that the user trusts his/her ears.
3. For IEMs, run some sine sweeps in the high-end, and ask the user if he hears some bump in the sweep. Narrow down the frequency range, until you find the one or two acoustic impedance peaks caused by the blockage of the ear canal, and compensate it until the user doesn't hear a bump anymore. For headphones, you're fucked. There's nothing you can do, especially without a 5128 database.

About the speaker calibration :

1. Pink noise is not very good for measuring speakers. Irregular magnitude response, less dynamic range, no phase measurement, and no windowing technique possible.
2. Because there is no windowing possible, you can't measure the direct sound of the speaker so you have to rely on third party Klippel NFS data. This is not ideal if there is a high amount of unit-to-unit variation on the model of speaker that is measured.
3. Ideally, you should compensate the whole listening window, not the on-axis response. This depends for every user. I know that IK Multimedia's ARC lets you choose different listening area sizes for this purpose.

So here's what you can do :

1. Use tried and tested sine-sweeps or MLS methods for the measurement of the speakers.
2. Calculate the direct sound using IR windowing.
3. Compensate the whole listening window by averaging the direct sound of multiple measurements taken in the listening area.

And one last thing, Dirac Live definitely adjusts its calibration for speakers working in unison. It's like the one thing that differentiates them from other solutions.

Hope this helps

 

Last edited: Mar 30, 2025

I like how some people from youtube and asr are meeting up here again.

 

1. There is a priority list behind the scenes for duplicate entries with Crinacle GRAS measurements being the highest priority for over-ears and super-review measurements being the highest priority for in-ears. The reason super-review was chosen was because his measurements nearly exactly match manufacturer GRAS measurements.
2. While the default reference track isn't the best, it is copyright free, broadband, spectrally dense, and has decent dynamic range. We have collected data on user's uploading their own song and compared it to results with the default track and the results are virtually identical.
3. Equalization after 10 kHz is performed, but not fine grain primarily due to limitations in the headphone performance and real world inconsistency with the predicted measurement result.

1. Not necessary since the only "clone" rig that is highly inaccurate is Crinacle's 711 coupler and it is the lowest priority in our database.
2. The blind element is critical for unbiased test results and is one of the main reasons the PEQdB targets are so good. I agree that the test track is not "ideal", but it would be best to use female pop rock
3. There is no real fix to poor acoustic performance in the treble. Moondrop Variations and Audeze LCD-5 are essentially acoustically optimal in higher frequencies and don't have issues there for this reason. Using sine sweeps to EQ won't get you there. The 5128 is also completely pointless.

1. We use periodic pink noise so this is isn't an issue since it has a flat magnitude response from the RTA point of view for every period. Phase measurements aren't necessary as long as the speakers are equidistant from the listening position.
2. Using windowing to correct the on-axis response instead of using spinoramas would be interesting for the reasons you mentioned.
3. The listening window calculated in spinoramas and the on axis response are typically very similar. Floyd Toole seems to use them interchangeably, but I wouldn't be opposed to basing the anechoic EQ on the calculated listening window response instead of the on-axis response.

So why are the left and right channels shown to be EQ'd to the same target?

Thanks for the suggestions, I really appreciate it! I will further consider the pros and cons of using multiple single point sweep measurements over periodic pink noise.

 

meanwhile Genelec tries to calibrate by using actual head scan and ears

 

OK, but what does "translates perfectly" actually means to you? Do you mean flat?

"coloration" is quite ambiguous, what do you really mean with coloration?

 

Genelec Aural ID is a very interesting tool, but PEQdB's headphone calibration focuses on tonality while preserving the "headphone sound."

The list of supported headphones is very limited and none are flawless like the Audeze LCD-5.

 

Last edited: Mar 30, 2025

"Flat" would be your true preferred low frequency response for a system combined with a flat on-axis/listening window response above the room transition frequency. Room resonances in bass, deviations from flat on-axis/listening window response above the transition frequency and directivity errors are colorations.

 

Mmm, I have a more simple approach, the bass must boom and pump and coloration is actually a good thing to me, I don't want to mix in an anechoic chamber.

Your ultra scientific approach doesn't actually meet my needs, that doesn't mean I'm right and you are wrong, it's just my way.

I support your effort and research, good luck, sure someone will jump on your boat.

 

Last edited: Mar 30, 2025

I don't support listening in an anechoic chamber, but low frequency equalization is essential for accurate bass