Re: Low latency filtering quality

[dynamicfusion]

Hello all. I have gone back and forth between "Linear Phase" and "Low Latency" DSP filter options the past few days and have come to the conclusion low latency adds an untuned piano string type of twang to the sound. Especially the upper frequency treble sounds. I have also had my fellow anan-operating neighbor switch between the two and I could call out in a blind test every time which filter mode he was using on his transmit. This occurs on both receive and transmit despite any buffer size applied.

I would encourage all of you to use Linear Phase if you want the highest quality sound input/output that the anan can provide. It is completely obvious there is a significant difference if you pay attention with decent headphones or speakers.

All the best.

linear phase.jpg (75.85 KiB) Viewed 6650 times

[W4WMT]

Hi Dynamicfusion,

I agree with your assessment of the sound of the minimum phase filters. The low latency is very nice, but it comes at a price. That said, the majority of hams, if not the vast majority, cannot hear the difference between the linear phase and minimum phase filters. In fact, I have had hams tell me I was just making it up

73 !

Bryan W4WMT

[W2PA]

The effect is quite noticeable on CW. It was discussed here.
https://apache-labs.com/community/viewtopic.php?f=9&t=3063&p=7296&hilit=w1ja#p7296

[w-u-2-o]

And yet CW is where low latency is most needed.

My subjective opinion is that most operators hear no difference, so making a blanket statement like that may not be helpful to all. This is based on many ANAN-to-ANAN contacts, including a daily rag chew net that is inhabited by a lot of ANANs.

Personally I hear no difference.

I'd like to see some objective proof of this effect, spectrograms or similar, to prove it exists.

If you are running linear phase filters, I would set the filter length to the minimum. Even the minimum length is already much sharper than any other filter in any other rig out there with the possible exception of the Flex Radios. That's the only setting that will come close to matching latency with the low latency filters.

[DL2XY]

And yet CW is where low latency is most needed.

My subjective opinion is that most operators hear no difference, so making a blanket statement like that may not be helpful to all. This is based on many ANAN-to-ANAN contacts, including a daily rag chew net that is inhabited by a lot of ANANs.

Personally I hear no difference.

You will not hear much difference if your RX path has LowLatency set. You always have these distortions and You get used to.

I'd like to see some objective proof of this effect, spectrograms or similar, to prove it exists.

Look at these step responses at -6 dB point of filter. These are made with onboard tools of Thetis (Testgen and Scope display).
Quality is not perfect, but you clearly can see a massive Gibbs effect.

LinearPhase 4096.jpg (110.04 KiB) Viewed 6575 times

LowLatency 4096.jpg (110.84 KiB) Viewed 6575 times

If you are running linear phase filters, I would set the filter length to the minimum. Even the minimum length is already much sharper than any other filter in any other rig out there with the possible exception of the Flex Radios. That's the only setting that will come close to matching latency with the low latency filters.

And this is also the setting for best quality audio!

cheers Walter

[w-u-2-o]

Interesting results, Walter. Would you please describe exactly how you did that? I tried to use the built-in transmit function generator to replicate that but could not accomplish it.

What I was able to do, however, was tune in a strong CW signal and look at it in scope mode with no APF or other processing involved. Filter was 400Hz, wide enough to show any ringing but not crazy wide for CW.

Some slight ringing with low latency at length 4K:

lowlatency4K.JPG (773.71 KiB) Viewed 6553 times

Far less ringing with low latency at length 1K:

lowlatency1k.JPG (640.05 KiB) Viewed 6553 times

No ringing with linear at 1K:

linear1k.JPG (751.54 KiB) Viewed 6553 times

No ringing with linear at 4K:

linear4k.JPG (684.83 KiB) Viewed 6553 times

Looking at this in an external audio analysis tool, and comparing the two polar opposite cases, linear 1K versus low latency 4K, again while the slight ringing is visible in the time domain, there's really no dramatic differences in the spectrogram.

lowlatency4k_1.JPG (343.04 KiB) Viewed 6553 times

linear1k_1.JPG (353.42 KiB) Viewed 6553 times

Again, I can't hear any difference, and the results at 1K are so little different spectrally only you golden ears can tell. I'd actually like to see a double blind test under controlled conditions, like the stereophiles do

[W2PA]

And yet CW is where low latency is most needed.

No argument, Scott. But I find that in using audio direct from the radio I have no noticeable latency on CW, which is quite important especially operating QSK. At reasonably high speed the audio effects with the low latency filter is objectionable enough to me to avoid using it.

With this setup, Thetis on the 8000 is the best CW rig I've ever used.

[dynamicfusion]

I can appreciate the low latency setting for cw, however my position is for best audio on voice, it's advantageous to keep linear phase. WU2O, I will do my own scope test and provide sound samples. Once I hear the twang, it's immediately blatant a fellow anan'r is using low latency mode. Will provide my findings later in the day.

[DL2XY]

Here are the settings to get the step response (time domain) with thetis.

To get a step response we must modulate the transmitter with a DC-signal in the form of a square
pulse.
This is done by setting the carrier frequency to 0 Hz (zerobeat). Set the testgen to pulse mode, a relative low pulse frequency (must match the scope timebase to get a standing waveform, it has no trigger) and a reasonable pulse with (Duty cycle) and rise time (Trans.).

Testgen.jpg (57.93 KiB) Viewed 6512 times

Next set TX and RX bandpass filters to 0HZ at the low end. A negative value would be better, but this is not possible in the current version anymore.
We are transmitting at zerobeat and the passband begins at zerobeat. That means our signal is at -6dB within the lower filter slope.

Switch all NR,NB,ANF,ANF,COMP and equalizers to off,DUP on, display mode to Scope.

Console.jpg (112.93 KiB) Viewed 6512 times

Set Scope timebase to 131us. This should give a slow moving waveform, if not modify to testgen pulse frequency to get a stable display.

display setting.jpg (65.34 KiB) Viewed 6512 times

Now you can play with the filter setting and activate mox.

dsp setting.jpg (62.4 KiB) Viewed 6512 times

[DL2XY]

This it how it looks at 16384

Linear Phase 16384.jpg (110.16 KiB) Viewed 6508 times

LoLat 16384.jpg (113.5 KiB) Viewed 6508 times

To set things right, these artifacs are not from a faulty implementation in Thetis.
Every filter does this - digital, analog and even mechanical filters.
They are pysically/mathematically properties of slope steepness and filter characteristics .

[dynamicfusion]

An excellent step by step! thank you for this.

[W1JA]

Some comments.

1. Low-latency filters add artifacts to the received audio that I can hear. For this reason I have switched to linear phase filters for all receive situations. For minimum latency I set Buffer Size to 64 (the minimum) and Filter Size to 1024. The filter size setting gives slightly wider filter skirts, but they're still better than most (all?) filters in other rigs.

2. It stands to reason that if the Linear Phase filters have better audio quality on receive, they have better audio quality on transmit. I have never tested this in an SSB QSO, but I've switched all my TX filtering to Linear Phase anyway.

3. Where I have tested TX audio quality is on FT8. I have noticed a significant improvement in the decodability of my FT8 signal when I use Linear Phase TX filtering. I get better S/N reports and more answers to my calls. I tested this repeatedly over a long time period and hundreds of QSOs. I know some others have also noticed this phenomenon.

4. Where low latency is most important is on CW QSK. I prefer to monitor using only my actual transmitted signal as received by Thetis (meaning that I turn off MON). I test QSK goodness (which includes latency) with my keyer set to 30 WPM and my external keyer's built-in sidetone turned on for the test. I send dits. The length of a 30 WPM dit is 40 ms. I can just barely hear a tiny timing difference between the keyer sidetone and my transmitted signal heard in the receiver. I perceive it as a very slight echo. (In actual operating on a noisy band, I often can't hear even this small echo.) Thus I conclude that RX latency is much less than 40 ms. I can send 30 wpm with my paddle, while monitoring my transmit only in the receiver, with no difficulty or confusion caused by latency.

73, John

[w-u-2-o]

Well life is all about options, and we are lucky enough to have software that is full of them. I'd still like to see a double blind test some time. The vast majority of users seem to hear no objectionable effects otherwise there would be a lot more chatter about this, both here and on the air.

Note this is not to say that linear phase is not the more "correct" approach, only to question whether it offers significant functional or audible improvements. Obviously you folks feel that way. To me it does not pass the "goofy test". It simply does not jibe with the effects imposed by the 800lb gorilla in the room, which remains the ionosphere. Once a perfectly pristine signal has left the antenna it will suffer group delay, phase distortion and doppler distortion far in excess of what the minimum phase "low latency" filters will do to it. Certainly that's not the case for QSK monitoring of a transmitted signal. But it is the case for signals coming in from thousands of kilometers away. By the time someone receives your signal they would not be able to tell what filter you used.

Some will claim they can, of course. To me this all devolves into the arguments that are endemic and timeless in the music industry, where sound is described in flowing, amorphous, non-scientific, i.e. artistic terms. At the end of the day it is sensation (outer truth) vs. perception (inner truth), and not everyone perceives the exact same stimulus in an identical fashion, and that sensation is 100% known to be colored by foreknowledge of the operating conditions, or some other set of preconceived assumptions about them.

[K1LSB]

Scott,

For whatever it may be worth, here are two recordings I've made of my transmitted audio, one using Linear Phase and the other using Low Latency. Please forgive the thoroughly unprofessional presentation, I was just ad-libbing both recordings as I went along.

I tried to explain my audio capture setup in the recordings. Please let me know if you (or anyone) have any questions.

Also please forgive the noisy 7000DLE fan in the background, I've recently moved and don't even have my shack set up yet, I've just got my ANAN sitting on a shelf at my left elbow and a very sensitive AT2035 mic in front of me.

I just wanted to put out an audio A-B comparison of the two filter types so anyone can listen to any differences. For my own part, I can't tell any difference at all audio-quality-wise between the two recordings.

https://drive.google.com/file/d/1-67lon ... sp=sharing

https://drive.google.com/file/d/1-7bO5T ... sp=sharing

Thanks,

Mark

[dynamicfusion]

What were your buffer sizes during this A/B test? Curious

[dynamicfusion]

This is my A/B at a tx buffer/filter sizes of 128/4096, which I normally operate at, comparing Linear Phase (test A) and Low Latency (test B) modes. The offending effect is not as obvious as at 64, but I can still pick apart the ringing twang left in the audio while Low Latency (test B) is active. This is from a receiver several hundred miles away.

It appears the larger the buffer/filter sizes in both columns are increased, the low-latency effect becomes less noticeable. Any way, as was said above, life with options is best. I'll stick with linear phase for now @ 128/4096. - Cheers

audio dsp.jpg (70 KiB) Viewed 6327 times

[K1LSB]

What were your buffer sizes during this A/B test? Curious

Not sure which buffers you're asking about so here are all of the relevant buffer fields I can think of:

VAC1 Settings.png (272.16 KiB) Viewed 6326 times

DSP Options.png (212.74 KiB) Viewed 6326 times

Hope that helps,

Mark

[w-u-2-o]

For whatever it may be worth, here are two recordings I've made of my transmitted audio...

Hi Mark,

Thanks for doing this. I can't tell the difference.

If you felt like doing more work, it would be better if you made a recording, then transmitted the recording with each configuration. Then both results could be put together as two tracks in something like Audacity (or some other DAW) and time aligned. That would allow someone to very easily switch back and forth between the two tracks during synchronized playback to try to detect any micrometric differences (BTW, do not do this as a left/right channel thing because most people do not have identical frequency response in each ear, you've got to listen with both ears, always).

Of course, if the differences are so small they require that kind of treatment to detect, then they are very unlikely to be heard as part of a normal signal under normal band conditions, as opposed the near perfect conditions you've obtained.

BREAK

It appears the larger the buffer/filter sizes in both columns are increased, the low-latency effect becomes less noticeable.

It is funny that you make this observation because it does not agree with the reality. It just another good example of how the perception of sound has just as much to do with expectations as it does with reality. Because the behavior here, if one was truly hearing a difference, should be the opposite of that. Consider that the linear phase filters have the same impulse response at any length, while the impulse response of the minimum phase, low latency filters gets more unruly with increasing length (increasingly steep skirts), just as Walter has shown above. So differences should be more apparent, not less, at increasing filter length.

BREAK

We tend to overdo things here during filter selection. How much more "brickwall" do we need? Consider the filter response plots taken from the WDSP Guide:

Capture.JPG (107.53 KiB) Viewed 6272 times

Capture2.JPG (94.21 KiB) Viewed 6272 times

You'll note that the BH7 response has much deeper stop-bands than the BH4 response, but less steep skirts. But that's easy to fix, just select the next length up filter and it's back to steep again. Before we had the minimum phase, low latency filters the latency penalty for going with a larger filter length was substantial. But with the min. phase filters it doesn't matter, latency is always the same from 1024 to 16384.

I'm just as guilty of overkill. For phone I run low latency BH7-4096. Before the low latency option was available I used to run BH4-2048. But I suspect I would be just as happy with 1024. And for CW, regardless of filter type, clearly 1024 is the way to go because even with a well behaved impulse response anything with a steeper skirt is going to treat the relatively square leading and trailing edges of CW signals poorly.

73,

Scott

[K1LSB]

For whatever it may be worth, here are two recordings I've made of my transmitted audio...

Hi Mark,

Thanks for doing this. I can't tell the difference.

If you felt like doing more work, it would be better if you made a recording, then transmitted the recording with each configuration. Then both results could be put together as two tracks in something like Audacity (or some other DAW) and time aligned. That would allow someone to very easily switch back and forth between the two tracks during synchronized playback to try to detect any micrometric differences (BTW, do not do this as a left/right channel thing because most people do not have identical frequency response in each ear, you've got to listen with both ears, always).

73,

Scott

That's a great idea, I hadn't thought of that (mostly because I'm not familiar with the audio recorder in VoiceMeeter). Lemme put a man on that and I'll post the resulting transmitted audio using each respective filter type.

Mark

[w9ac]

I just took a series of measurements that compare QSK CW latency as a function of filter type and filter size. There's a significant latency difference between low latency and linear phase filter types. Tests were performed at my ANAN 7000's headphone jack, and ASIO audio from a Presonus Mobile 192 sound interface. In each measurement, a buffer size of 64 was used.

When using a low latency filter, there's essentially no latency difference between a filter size of 1024 or 16384. When selecting linear phase, latency goes from about 35 ms with a 1024 filter size to 220 ms with 16384.

I can post oscilloscope screen shots if there's an interest, but with so many screen captures, they tend to clog the discussion. Obviously low latency is the filter type of choice when using QSK.

Between ASIO and the headphone jack, ASIO latency is slightly better. However, as I posted a long time back, latency from the headphone jack builds over time. It's been that way with each Thetis version going back to its beginnings. Two other ops have confirmed the same result. So, for this reason, I prefer listening to ASIO audio. Not that it matters, but it also provides the cleanest audio path between the transceiver and my ears.

Paul, W9AC

[K1LSB]

You'll note that the BH7 response has much deeper stop-bands than the BH4 response, but less steep skirts.

Scott,

I did notice that early on. Given that I operate exclusively in phone mode, I decided I'd rather have the steeper BH-4 skirts so as to pose the least interference to any QSO that happens to be immediately adjacent to my passband. Besides, even the relatively shallow stop-band of BH-4 is still much deeper than even the best IMD cleanup that PureSignal is capable of applying to my transmitted signal, so the residual splatter of my signal is still at least 30dB greater than the BH-4 stop-band. Effectively, the BH-4 stop-band is of no consequence during phone operation, so I'll gladly take advantage of the steeper cutoff.

Mark

[w-u-2-o]

I just took a series of measurements that compare QSK CW latency as a function of filter type and filter size. There's a significant latency difference between low latency and linear phase filter types. Tests were performed at my ANAN 7000's headphone jack, and ASIO audio from a Presonus Mobile 192 sound interface. In each measurement, a buffer size of 64 was used.

Paul,

You are bringing back memories. This was done to death by myself and Bryan, W4WMT, during a period that predates this forum by some time. It was a "war on latency", and it ultimately lead to Warren coming up with the minimum phase, low latency filters, and a few other very important changes to buffering of the IF data. I had originally proposed IIR filters, but since everything in WDSP was FIR based, Warren was more comfortable maintaining an FIR approach, and he said it made it much easier for on-the-fly filter construction (which how we are able to drag passband edges around in real-time). However, had he done an IIR implementation, latency would be much nearer to zero!

When using a low latency filter, there's essentially no latency difference between a filter size of 1024 or 16384. When selecting linear phase, latency goes from about 35 ms with a 1024 filter size to 220 ms with 16384.

You might want to go back and recheck those numbers. Where the linear phase FIR filters are concerned, the length of the delay through the filter is easily calculated as length/sample rate, the latter being 48KHz. Note that this is reference to what is essentially the 2nd IF sample rate, the passband sample rate, not the audio sample rate, although they are identical.

Hence the delay for a 1024 long filter is 21.3ms, and for 16384 long filter is 341.3ms. To that you have to add the delays associated with various buffering throughout the firmware and the software. It ought to be possible to deterministically know what that additional delay should equal, but nobody has ever sat down to figure it out, other than to know the entire delay path is somewhere in the 30 to 40ms range, much as you measured, with a length of 1024. The same is true for the low latency filter, but it does not get longer with increasing filter length, which is the entire point of the low latency filter.

Between ASIO and the headphone jack, ASIO latency is slightly better.

Again, we studied this incessantly during "the war on latency". It helps that going to from ASIO avoids additional trips up and down the Ethernet cable. All audio flows through Thetis, the ANAN hardware simply includes what amounts to a custom sound card interface. IMHO, the original architecture should never have included that.

Don't forget that you've got latency in two directions. RX latency and TX latency are both important. QSK is an obvious example. But it's true on phone, as well. If you have not done a good job at tuning the system, and have all the audio bells and whistles engaged, TX latency can get upwards of 300ms. This makes RX-to-TX overall delay even longer. It starts to become hard to break into ragchews or, worse, you can forget about trying to do well in a phone contest. With the advent of the low latency filter it became possible to turn all of those audio bells and whistles on again without driving delays into the stratosphere. Every feature you use, from the leveler to the CFC, exacts a delay. You could have a bit of fun with your scope seeing just how much

73,

Scott

[w9ac]

Scott,

Here are two of today's ASIO output screen captures. Triggering is accomplished with a modified Winkeyer. It has a dedicated scope output that when keyed goes from +5V to 0V for the duration of the keyed element. The scope is set to trigger on the leading edge of the transition. A single "dit" is shown. Scope is set to 20 ms/div.

16384 linear phase filter size delay is closer to about 210 ms. Close enough although I could have placed markers at the trigger and beginning of the QSK sidetone waveform to get an accurate result. The goal was to see the effect of incremental latency as a function of filter size, not a highly accurate measurement of exact delay time.

Paul, W9AC

[w9ac]

And here's an ASIO screen capture using the low latency filter type and 1024 versus 16384 filter sizes. Essentially no latency change. CW QSK is fast and smooth with any filter size.

Filter bandwidth in each case is 400 Hz. I neglected to try it at extremely low bandwidths like 25 Hz.

Paul, W9AC

[w-u-2-o]

Paul, just to clarify: the blue trace denotes keying of some other radio, the yellow trace is the audio output of the ANAN. Is that correct?

[w9ac]

Paul, just to clarify: the blue trace denotes keying of some other radio, the yellow trace is the audio output of the ANAN. Is that correct?

No other radio is used in the measurements.

The yellow trace is ANAN audio at the Presonus Mobile 192 main output jack.

The blue trace is a dedicated open-collector from a modified Winkeyer. Inside the Winkeyer, the NPN collector is pulled high but goes low on each keyed element. It's used only for scope triggering. It's split from the same logic line that keys the ANAN from the Winkeyer's KEY OUT jack. That jack connects to the ANAN's KEY IN jack on the front panel. Of course, a paddle is connected to the Winkeyer and not directly to the ANAN.

Paul, W9AC

[W4WMT]

Good! So for the 1024 minimum phase example, about 16ms worth is the receive latency and the rest is on the transmit side. Very close to what I am measuring here.

73,
Bryan W4WMT

[K1LSB]

Scott,

Per your suggestion I've made an audio recording using the VM Potato recorder and transmitted that using both filter types (Low Latency and Linear Phase) on 3.925 MHz into a dummy load, capturing a separate recording of both of those transmissions using the audio recorder in SDR Console.

I was able to achieve a quieter shack environment this time by making sure my 7000DLE was turned off prior to recording the original audio in VM.

Here are those recordings, I can hear no difference but my ears are shot.

Linear Phase: https://drive.google.com/file/d/1EovwGA ... sp=sharing

Low Latency: https://drive.google.com/file/d/1O8YJjy ... sp=sharing

Thanks,
Mark

[w-u-2-o]

Great job, Mark!

I put both files into Audacity as separate tracks. I trimmed them so that they will playback in sync to within 0.5ms. I had to trim both because I over-trimmed the first one by 2.5mS, so had to chop 2.5ms out of the second one. That trimmed files are attached here. It's easy to do this in Audacity, just zoom in until the actual waveform is clear and match them across both files, and trim away at the front until things line up.

But I've done this for you. So drag and drop each trimmed file attached below into Audacity and they'll go in as two separate stereo tracks. Then hit the play button, then you can use the "solo" buttons to jump back and forth between the tracks while they play back.

Damn if I can hear any difference. If you zoom in far enough you can see small differences in the waveform but I can't hear them. Those differences are probably because of the compressed format (wma) that you chose. To really do this 100% correctly this would need to be done using a lossless format (e.g. wav), but then file sizes get very large. Channel noise on HF will be many times worse than any distortion the lossy compression might cause.

Thanks!

Scott

[K1LSB]

Thanks Scott. I could certainly look into whether SDR Console offers a lossless audio recording format and re-record a couple of transmissions this evening if you wish (I'm at work right now).

At any rate, it appears you've answered the question of whether the LowLatency filter has any discernible effect on transmitted phone audio. It appears that it does not (at least with the DSP options I'm currently using).

Thank you very much!

Mark