rhasspy works with raspberry pi zero w?
Not really, it can with basic functions, pi3A+ or above is prob recommended.
I presume some will shout I have it working but we all have different opinions on the worth.
Pi zero has become an if for me now and used to be a fan but so much now is annoyingly just short and if it had 2 core, if it had a neon it would be great but prob best spending on a pi3A+ as will give you x10 more and its a lot of return for that $10 more.
It does technically work as a Rhasspy satellite but, as @rolyan_trauts said, the performance is quite poor for the money.
In addition, the ARM CPU it uses is different enough from the other Pi’s that it requires special handling when building all of the Docker images and Debian packages. I would like to eventually drop “full” support for the Pi Zero, and have only a small satellite package that just streams audio.
I think the Zero’s time might be coming to an end especially for AI.
The new ESP32-S3 have vector acceleration and likely outperform in cost and function and for a project like this maybe the last nail for the zero.
How much is sales pitch and truth we will find out but you can count it as gospel that KWS will run as currently the current ESP32 can manage simple models that the vector accelerated LX7 MCU of the s3 should do with relative ease.
Price is likely to end up like the ESP32 and doesn’t need a SD card.
All the KWS should be KW and stream till silence or notification and be completely free of proprietary protocol such as Hermes audio and a intermediary KWS server should then use that protocol if there is wish.
rhasspy-microphone-cli-hermes needs a rethink as firing uncompressed wavs over MQTT that is encrypted is a complete load (load being the prime word) whilst not supporting codec or compression.
The is no need for a permanently broadcasting microphone and one that only broadcasts from KW to end of command sentence does not need encryption.
Really it should send MFCC as its not audio that is required and without compression it can be up x16 smaller than the wav data and x32 if used with streaming compression.
If someone can break your network wifi security or snoop onto a cable then create a MFCC to voice synthesizer run on models captured from the users voice, they have managed to garner you tell Rhasspy to ‘turn off the lights’ from time to time.
No-one is and if they did who cares?
I have Rhasspy running on a Raspberry Pi Zero WH with Respeaker-2-Hat.
But only as a Satellite + Hermes LED Control.
It is running constantly at 95%-100% CPU and the reaction of the Pi ist ok, but not more.
If you plan to do more than this with the Pi (squeezelite, etc.), you will not end up happy.
I am currently investigating a Rock Pi S, but there are other options like the Banana Pi Zero or Orange Pi or Nano Pi, etc. They all use a kind of quadcore CPU.
But this is only my experience.
There is actually a lot you can do on a Pi Zero considering if you not constantly doing a whole lot of pointless and you use load wisely in a state machine of providing just the action that is needed at anyone time.
I have never run Rhasspy-satelite as I have never seen such tidy pythonic code of a totally pointless implementation of Hermes audio thats only purpose is to encapsulate an audio stream and give it a branding.
I have recently been reading Mulitple Satellites and Multiple Masters of an extremely competent user battle for 2 weeks to get multiple satellites to work which they did but the end result is that the satellites work on a first serve method so if your wired satellite that is at the other side of the room than your wifi satellite next to you always transmits the far worse signal because its always first in!
I have no doubt that your Pi Zero is running at 95%-100% CPU whilst dubiously constantly broadcasting a mic to chunk up and broadcast over encrypted MQTT for no other reason than they could.
You do not need a quadcore to do that or at least you shouldn’t have to if Hermes audio and Rhasspy satelite wasn’t such a huge brainfart.
PS the RK3308 great little A35 board with hardware VAD but did you ever get the analogue mics to work without a whole load of noise?