Lowest possible latency

[robiwan]

I've tried finding the latency of the WM8731 codec, but there's no specific information on this in the datasheet. I suspect it to be low though, on the order of < 0.5 ms.

I've bought the soundcard in order to make a system with as low a latency as can be achieved. I suspect a 32 sample ALSA period size (~ 0.7 ms @ 48Khz) should be achievable on a R Pi 3 ?

Any feelings/thoughts on this ?

[robiwan]

To clarify, I'm seeking to minimize the latency of the audiopath:

ADC -> ALSA -> DAC

[flatmax]

I have run the ALSA utility latency to find it is of the order of 500 us (half a microsecond). This was found using very low CPU load.

Check here for example :
https://www.kickstarter.com/projects/12 ... ts/1606640

I would start with 2 periods of 1ms and see if that works nicely. If so, then reduce the sample count accordingly.

Can you post your findings back here please ?

thanks
Matt

[robiwan]

Thanks! Will do so.

[robiwan]

Ok, I've setup my Pi with the stereo board (and have been streaming Spotify via Mopidy, sounds really good! )

Anyway, built the ALSA latency test and got following results:
./latency -r 96000 -f S32_LE -p
Maximum read: 56 frames
Maximum read latency: 583.333us, 0.583333ms (1714.2857Hz)
Playback time = 131878.455664, Record time = 131878.455662, diff = 2

./latency -r 48000 -f S32_LE -p
Maximum read: 48 frames
Maximum read latency: 1000.000us, 1.000000ms (1000.0000Hz)
Playback time = 132038.521231, Record time = 132038.521228, diff = 3

Not bad at all. The application I have in mind will do a lot of FFT convolution, so that's an aspect also, to have enough time to compute everything during a callback period.

[flatmax]

OK - sounds like you have a real time application on your hands.

How many Fourier bins are you targeting ?
I find that the more Fourier bins I use, the lower the latency requirement.

Matt

[robiwan]

Depends, I'm doing FIR filters in the frequency domain, so it can be 4 instances with lengths of up to 0.5 s (24000 taps @48K). I'm planning on a partitioned implementation to remove the latency due to FFTs. Maybe dual-rate to lower the lengths of the IRs. I'll see how it goes. Also, it is possible to use the GPU for the FFTs, which should heighten performance significantly.

[flatmax]

Oh be interested to hear if you get GPU computation for DFTs going !

Sounds like you have nice long filters. Are you running FIR eqs or a different application ?
You writing the software yourself ?

Matt

[robiwan]

It is for a room correction application, so 0.3 to 0.5 s length of FIRs usually are adequate, and yes, I'll be writing the software myself. Right now I'm trying to determine where to apply the processing. Since I want to do both in/out low latency and be able to stream audio (like Spotify), I think an ALSA PCM plugin would be the appropriate place.

[flatmax]

Room correction - nice one.

I agree, an ALSA plugin would probably be the best. That way you can set it to be default and not even have to think about it.

I have an ALSA Plugin class as part of gtkiostream. An example is here :
http://gtkiostream.flatmax.org/ALSAExte ... ample.html

You can also interface Eigen matrices to that, which come with an fftw inferface. You may find that useful.

Are you planning to use C/C++ ?

Matt