Thats the truth as have been trying to tell you.
Coming back to the initial topic 
This has definitely to do with pyAudio specifically.
On 1 of my PI, I reverted to arecord on Friday and no crash for the last 3 days. On 1 of my PI still running pyAudio, I got this input overflow in 1.5 hour after docker restart this morning…
That doesn’t surprise me. I remember seeing similar messages when I was testing hermes-audio-server, which uses PyAudio, and I never found out how to prevent these frome happening. It was quite erratic and I spent a lot of time trying to debug these messages. But now with the ReSpeaker hardware I don’t have this problem anymore with Rhasspy 2.5.0-pre.
Which Respeaker HW do you use?
I have 2-mics PiHat and 4-mic Array (the square one) but the arecord output has much less background noise with Jabra 410 compared to both Respeaker devices. I guess it should be better for Wakeword and ASR, that’s why I stick to the Jabra for all my PIs (which was my Snips setup, btw).
I’m using the 2-Mics HAT and I also have the 4-Mic Array lying here, but haven’t found the time yet to test the latter.
Let me followup on this. After switching to arecord (instead of Pyaudio), the issue has fully disappeared. Therefore I will close the topic as Soved.