Thanks for the reply.
Not off to a great start I’m afraid. Downloaded the recommended image (recommended_mark_2.img.xz) and wrote it to an external SSD - the system hung at a systemd startup script.
Next I tried writing the image to the original Dinkum drive and performance was somehow worse, it takes about 10 minutes to boot and takes about 15-20 seconds to respond to “hey neon”
So, I guess the quality of the flash drive is critical. Is there any way I can make this thing work with an SSD?
You’re absolutely right, the drive quality is a critical factor. SSDs are the best way to go. If you’ve got one, just write the image and plug it in like a USB, then plug in power. If you don’t have one hit me up for a coupon, or buy one anywhere you like. It doesn’t take a high-end SSD to get excellent performance.
I’ve managed to get an SSD that works with it - an older Kingston 120GB SSD. I guess the first SSD I tried the Mar-II didn’t like the enclosure.
Anyway, performance on this thing is still pretty bad. I’ve recorded a quick video to demonstrate some of the issues. While I was running the test load average peaked at about 9, which on a quad core CPU is not good.
In the video you can see where the performance gets so bad Neon starts to miss words - this is the most problematic part as it means I need to repeat myself sometimes 3 or 4 times.
Is the kind of performance I’m seeing in the video typical?
I gave up in the video, but Home Assistant did finally open about 5 minutes later.
Thank you, video was very helpful! I’ve made a quick one in return, with near-zero production value, but perhaps significant utility. 
A couple things -
First, except for when you are a responding to a query from Neon, you don’t need to wait for the “boop” sound to start talking.
Our most recently released stable version (and of course the current alphas) include a looped listening function, which has Neon go back into memory after recognizing the wakeword, and retrieve the audio that followed the wakeword. I’m at a loss of how to make that apparent to users, suggestions welcome. I’ll certainly explain it in our upcoming user newsletter & posts.
Second, that’s not terrible speed, but it can do better with either a faster SSD or upgrading the Raspberry Pi from 2GB to 4 or even 8. We’re also continuing to make speed improvements on the code side.
It will occasionally miss words or call a skill incorrectly, but not so frequently. I’m not sure what is going on with that. If you’d like to have your machine “create a troubleshooting package” and send it over to me with a quick note so that I know it’s from you, we can take a look at what’s been going on inside.
Here’s my quick and very much not for official publication for obvious reasons video:
I did the same 3 queries, with a “good” but not high-end USB, using our stable version & then our most recent alpha.
If anyone else would like to make some good videos for us (tripod, no unwanted reflections, etc 
), please do message me. We’d be happy to facilitate that.