10: Latency, C++ Real-Time Audio Programming with Bela

Lecture 10 of C++ Real-Time Audio Programming with Bela. This lecture explains latency, where it comes from and how to measure it, showing Bela's 1ms latency in action.
In this lecture:
Section 1: Explaining latency 0:00
Section 2: Sources of latency 8:41
Section 3: Measuring latency 17:26
Section 4: Automatic latency tester 24:25
Example code and companion materials:
github.com/BelaPlatform/bela-...
This course is a deep dive into how Bela processes data, and how to implement real-time audio and interaction. If you want to learn or improve your skills with C++ and audio programming, this is a great way to start!
Bela is the open-source platform for creating beautiful interaction. If you’d like to get a Bela system to follow along with these lectures, visit shop.bela.io. Use promo code CREATEATHOME for £10 off Bela and Bela Mini Starter Kits.
Music by Vula Viel (vulaviel.com).

[@kostantinosmag2986]

Great ....Nice one....

[@maehw4070]

Why is the latency measured at two differenct edges? The rising edge on channel #1 and the falling edge on channel #2?

[@apm414]

Good question. The way the Bela audio codec is set up, the passthrough produces a phase inversion, so the rising edge on the input corresponds to the falling edge on the output. I perhaps should have mentioned that in the video! This could be eliminated by multiplying the output by -1 before sending it, but I chose to use the built-in passthrough example.

[@DmitryIvanovDfcreative]

I wonder if there was another sort of mic and ADC, structurally similar to cochlea with frequency domain data as output. It would populate input buffer immediately with 0 latency (digits there would always be relevant). Processing as well would be simpler for many cases, like filtering. Outputting is potentially also 0 delay. It looks like the whole set of latency problems is caused by initially picked time domain.

[@joeblack375]

The cochlea essentially performs a frequency-to-location transform due to its continuously varying impedance. You would still have to determine/digitize the level in multiple locations. If all you need is a continuously updated FFT (instead of a cheap, block-wise computation), there are algorithms for that. Processing completely in the frequency domain would in fact result in *infinite* latency, as you have to know the entire signal a priori. In any case, there *are* cochlea-like chips (for HF processing): dspace.mit.edu/bitstream/handle/1721.1/59982/Mandal-2009-A%20Bio-Inspired%20Active%20Radio-Frequency%20Silicon%20Cochlea.pdf;sequence=2

[@joeblack375]

The data sheets show 21 and 17 samples of (filter) latency, the measured hardware latency (both on the Bela and the oscilloscope) is 41 samples, leaving three samples unaccounted for. Would be nice if you explained those as well (fairly sure two of them are due to serial transfer via I2S which obviously requires time separate from the built-in anti-aliasing filter).

[@apm414]

Good question. I would expect 1 frame of latency on each of the input and output from the I2S transfer, as you say. The remaining sample of latency is probably the effect of slow rise time of the output signal because of the band-limited audio output. If you turn down the overall level of the output in the Bela settings and run the same code, you'll get another 1-2 samples of apparent latency because of the time it takes the input to cross the threshold specified in the code.

[@richardsewell5159]

I'm interested in measuring the latency between USB midi controller pads (MPD218) and Bela output. Best way I can think of is to set a microphone over the pads and plug that to the left channel of my DAW and then record bela output on the right channel and look at the difference between pad press and bela output. The midi callback routine is an auxiliary (non-realtime) task on Bela, so I'm not sure what to expect, but it seems quite tight at the moment. I'm considering hooking up the MPC pads directly to ADC inputs on Bela, but not sure if it's really worth it. p.s. I want to hear a cheer in this video when you measure 1ms latency ;)

[@oromoiluig]

yes, that should work. It's the same approach that was used here www.tandfonline.com/doi/pdf/10.1080/09298215.2017.1340485?needAccess=true