Sound Quality in release 4.1

[rhizomusicosmos]

Concepts such as "musical" and "veil" when applied to audio are extremely difficult to quantify and therefore troubleshoot. These subjective terms have as much to do with an individual's personal context and cognition as with anything tangible. That's not to say there aren't quantifiable triggers to these perceptions, e.g. distortion levels, sampling artifacts, EQ differences or even just a difference in overall level.

Skip is suggesting that the mid-range quality is different -- too much 'presence'. If there is a convolutional difference then this can be measured and I might be able to do this. I think I still have an SD Card with 3.8.4 on it, so I could do a comparison.

However, this won't identify what has caused the changes even if we can find a measurable difference . . .

What do people prefer: the RT or LL kernel?

-Richard

[DRONE7]

Dick Shahinian...(he of the wondrous Shahinian Loudspeakers...) once told me that " equipment can often be like an old shoe...you've worn it a while and it fits...when it was new you thought it didn't fit and then you wore it a while..." :-)

http://shahinianacoustics.com/home.htm

[RafaPolit]

Concepts such as "musical" and "veil" when applied to audio are extremely difficult to quantify and therefore troubleshoot. These subjective terms have as much to do with an individual's personal context and cognition as with anything tangible. That's not to say there aren't quantifiable triggers to these perceptions, e.g. distortion levels, sampling artifacts, EQ differences or even just a difference in overall level.

Skip is suggesting that the mid-range quality is different -- too much 'presence'. If there is a convolutional difference then this can be measured and I might be able to do this. I think I still have an SD Card with 3.8.4 on it, so I could do a comparison.

However, this won't identify what has caused the changes even if we can find a measurable difference . . .

What do people prefer: the RT or LL kernel?

-Richard

My experience was that LL was better.  From what I recall, people using external USB DACs tended to prefer the LL one, while people using internal i2s solutions overwhelmingly favored the RT Kernel.

Again, of course, just subjective appreciations and what I remember reading, so take it with a pinch of salt.

Best regards,
Rafa.

[ukbrown]

In my case, I cannot state that 3.8.4-LL is better than 4.1 with the same settings elsewhere as a general statement. It does sound better in my system, but my system is built for my old ears and preferences and I have no reason to believe that other folks with their ears and their systems wouldn't prefer 4.1.

For me, the 3.8.4 sound has a different tonal balance versus 4.1. I hear no true problems like distortion, just a little too much 'presence' that suggests to me that the midrange is a bit strong. Perfect equalization could well do the trick. I have tried the equalization before and I could really hear some degradation. I'll try it this evening with 4.1 to see if the earlier degradation is still there.

Skip

I did not notice any real difference in sound.  Most of my listening is done through an ak70 and Shure headphones and the pi and my budget Marantz amp don't sound much different. It still sound better than my CD player

[rhizomusicosmos]

OK, I've just started the process of testing but this is my method:

Created WAV test files in Sound Forge (all at -6 dB FS):
20 Hz - 48 kHz tone sweep @ 24 bit 96 kHz;
20 Hz - 44.1 kHz tone sweep @ 24 bit 88.2 kHz;
100 Hz sawtooth wave @ 24-96;
pink noise @ 24-96.

Chose some test music (all in WAV @ 16 bit 44.1 kHz):
Toto - Rosanna (musical equivalent of pink noise)
Alan Parsons - Prime Time (hi-hat sounds and other instruments are staggered and well mixed)
Stravinsky - L'oiseau De Feu - Xiii. Danse Infernale De Tous Les Sujets De Kachtcheï by Boulez/CSO (dynamic orchestral)
Jordi Savall - Turquie, mss. de Kantemiroglu - Makam Rast Murass'a usul Düyek - from the album Orient Occident (good room and instrument sound)
David Sylvian - When Poets Dreamed of Angels - from remaster of Secrets of the Beehive (problematic treble on some systems)

I loaded these onto a USB stick for replay in an RPi3 running moOde 3.8.4 and 4.1. I am using a good SMPS with plenty of amps for power.

DACs:
HifiBerry DAC+ Pro IIS (I don't have an ALLO Boss, sorry, but this does use the PCM5122 chip)
Topping D10 USB
Topping D30 USB

I am recording the output of these DACs with an RME Fireface UFX and Sound Forge Pro software at 24 bit 96 kHz.

I will switch between LL, RT and Standard kernels in moOde 3.8.4 and record the output of the test files. I will do the same for moOde 4.1 with its standard kernel. I will then use FFT analysis to find differences or anomalies. I hope that this will show any convolutional changes between the kernels.

I will also spend a bit of time listening to music, both normal and hi-res. I will be listening on Genelec 8260A speakers and Beyerdynamic DT-880 Pro headphones. The connections are pseudo-balanced from the DACs to an Allen & Heath mixer (RCA -> TRS) and fully balanced from the mixer to the Genelecs (XLR -> XLR). The listening room is acoustically treated with a very short RT60 and quiescent noise level of 25 dBA. The speakers have been calibrated for the listening position with the Genelec SAM system.

Caveats:
This will take a little bit of time . . .
This is not a laboratory-standard test. I don't have an Audio Precision but if something interesting pops up I can look at using an analyzer.
I am hoping the RME Fireface and other equipment are of sufficient quality to show audible differences.

I suspect some of the perceived differences are psychoacoustic and I can't measure for those, but I am going into this with an open mind.

If you have any suggestions about kernel settings or testing methods, etc, I am very happy to try them out. For example, I'm thinking that a square wave (odd harmonics) might be better than a sawtooth (all harmonics).

-Richard

[philrandal]

Concepts such as "musical" and "veil" when applied to audio are extremely difficult to quantify and therefore troubleshoot. These subjective terms have as much to do with an individual's personal context and cognition as with anything tangible. That's not to say there aren't quantifiable triggers to these perceptions, e.g. distortion levels, sampling artifacts, EQ differences or even just a difference in overall level.

Skip is suggesting that the mid-range quality is different -- too much 'presence'. If there is a convolutional difference then this can be measured and I might be able to do this. I think I still have an SD Card with 3.8.4 on it, so I could do a comparison.

However, this won't identify what has caused the changes even if we can find a measurable difference . . .

What do people prefer: the RT or LL kernel?

-Richard

My experience was that LL was better.  From what I recall, people using external USB DACs tended to prefer the LL one, while people using internal i2s solutions overwhelmingly favored the RT Kernel.

Again, of course, just subjective appreciations and what I remember reading, so take it with a pinch of salt.

Best regards,
Rafa.

My experience on a Pi Zero W with I2S DAC (Pimoroni pHAT DAC) was that the RT kernel sounded best in 3.8.4.  LL sounded harsh to my ears.

In the early 4.0 betas, I found the RT kernel unusable, for whatever reason (kept getting audio stutters).

Phil

[TheOldPresbyope]

 . . .

What do people prefer: the RT or LL kernel?

-Richard

My experience was that LL was better.  From what I recall, people using external USB DACs tended to prefer the LL one, while people using internal i2s solutions overwhelmingly favored the RT Kernel.

Again, of course, just subjective appreciations and what I remember reading, so take it with a pinch of salt.

Best regards,
Rafa.

My experience on a Pi Zero W with I2S DAC (Pimoroni pHAT DAC) was that the RT kernel sounded best in 3.8.4.  LL sounded harsh to my ears.

In the early 4.0 betas, I found the RT kernel unusable, for whatever reason (kept getting audio stutters).

Phil

There may be a causal explanation for this difference in opinion between USB- and I2S-DAC users concerning LL- vs RT-kernels. 

I commend the series of blog entries at https://LeMaRiva.com about Preempt-RT built against kernel 4.14.y on RPi3B and RPi3B+ (and, yes, LeMaRiva gives instructions for building Preempt-RT on 4.14.y kernels; the patches are now a branch of the official Raspberry Pi kernel repo).

Here's a revealing quote from the end of #Raspberry Pi: Preempt-RT vs. Standard Kernel 4.14.y

Performance! (updated 26.02.2018)

I found a problem on the patched kernel, and I thing the problem is in every Preempt-RT Raspbian versions. The IRQ/39-dwc_otg process uses more than 30% of the CPU! That is too much! The interruption is related with the USB irq: A known problem of the Preempt-RT patches.

This should be a reminder that there is more in play here than simply substituting one kernel for another to lower kernel latency. It can affect a number of aspects of the system---temperature, USB, network, etc., and not always kindly.

Conclusions:

The Preempt-RT patched Raspbian kernel (4.14.y-rt) offers a solution to reduce the kernel latency (see results here). But, you lose a lot of CPU and communication performance. The data transfer over Ethernet is reduced to 34% and the CPU performance up to 12%.

LeMaRiva has some observations about temperature-throttling of cpu speeds which are relevant if one is packaging an RPi3B/B+ in close quarters.

His blog entries cover a variety of topics, are interesting, and well written. Give him a read.

Regards,
Kent

[RafaPolit]

...

If you have any suggestions about kernel settings or testing methods, etc, I am very happy to try them out. For example, I'm thinking that a square wave (odd harmonics) might be better than a sawtooth (all harmonics).

-Richard

This sounds great! We will surely learn something from this, thank you.

The one suggestion I have is that for all these tests, you have any resampling dissabled within the MPD options.  Just to take away other variables out of the equation, otherwise we don't know if you are measuring the different MPD resampling implementations or the kernel.

Same goes for volume leveling, equalizers, etc. that may have an effect on sound.  Make it as bit perfect as possible (which I think is 100% in this case).

Thanks!
Rafa.

[rhizomusicosmos]

There may be a causal explanation for this difference in opinion between USB- and I2S-DAC users concerning LL- vs RT-kernels. 

I commend the series of blog entries at https://LeMaRiva.com about Preempt-RT built against kernel 4.14.y on RPi3B and RPi3B+ (and, yes, LeMaRiva gives instructions for building Preempt-RT on 4.14.y kernels; the patches are now a branch of the official Raspberry Pi kernel repo).

Here's a revealing quote from the end of #Raspberry Pi: Preempt-RT vs. Standard Kernel 4.14.y

Performance! (updated 26.02.2018)

I found a problem on the patched kernel, and I thing the problem is in every Preempt-RT Raspbian versions. The IRQ/39-dwc_otg process uses more than 30% of the CPU! That is too much! The interruption is related with the USB irq: A known problem of the Preempt-RT patches.

This should be a reminder that there is more in play here than simply substituting one kernel for another to lower kernel latency. It can affect a number of aspects of the system---temperature, USB, network, etc., and not always kindly.

Conclusions:

The Preempt-RT patched Raspbian kernel (4.14.y-rt) offers a solution to reduce the kernel latency (see results here). But, you lose a lot of CPU and communication performance. The data transfer over Ethernet is reduced to 34% and the CPU performance up to 12%.

LeMaRiva has some observations about temperature-throttling of cpu speeds which are relevant if one is packaging an RPi3B/B+ in close quarters.

His blog entries cover a variety of topics, are interesting, and well written. Give him a read.

Regards,
Kent

So according to that info, it appears that the RT kernel may be stressing the system and possibly affecting USB and ethernet performance. I'll check the system stats while running the tests and see what difference there is.

Thanks for that, Kent.

-Richard

[Kisszabo]

Thank you for your work - i follow up!