[FYI] new Resonate protocol

[luar123]

Just wanted to let you know that that esphome and music assistant devs are working on a new sync protocol as an alternative to snapcast. The timing algorithm seems similar but overall the protocol offers more flexibility and functionailty and should be better suited for low power devices. Interesting is that they are using a kalman filter instead of the median filter for the time offset.

Here are some resources:
https://github.com/Resonate-Protocol
https://github.com/orgs/music-assistant/discussions/3883
esphome/esphome#10183
https://discord.com/channels/753947050995089438/1369735163629801532

KevinA
Luckily math is here to save the day 😆 . A Kalman filter is a nice way to combine measurements with different uncertainties, which is exactly what we have! Basically, if the most recent message has a very low delay, then we weigh its value much more than if it had a very high delay. It gives us a much more stable offset than using the median, but it also gives us an estimated error after each update. This means the client can estimate how accurate its own offset is! This would allow us to slow down or increase the pace we send the time messages. The more confident we are in the offset, the less we need to send - and vice versa. This will be especially helpful when we first connect and need to spam a bunch of time messages to quickly build our confidence in the computed offset.

So being accurate is important, of course, but being stable is also very important. The median was problematic, even if its inaccuracy wasn't audible. With the median filter, the offset would shift by over 200 microseconds, on average, per update, which means the client would have to constantly be tweaking the audio to try to stay in sync. With very big jumps, it would require hard syncing which is audible.

The Kalman filter modeling the drift (of 2 seconds per day) had an average offset change of only ~25 microseconds per update. Keep in mind, if the clock does drift 2 seconds per day, we would expect it change 23 microseconds per update if the updates are 1 second apart. If there is no systematic drift in the client clock, the Kalman filter had an average change of less than 1 microsecond per update.

[CarlosDerSeher]

Sounds promising, maybe it will be possible to implement this too. Thank you

[Hedda]

FYI, I also collected links and information about this "Resonate protocol" here in Music Assistant discussions section for general discussion:

• https://github.com/orgs/music-assistant/discussions/4200