Just updated Whisper to 3.0.1 and its now using 21.7% of the system RAM out a a total system wide usage of 37.8%, seems somewhat excessive in that it’s using at least another 1Gb of memory compared to the previous version. (1.74Gb in total…).
Or is it just me ? (eg. vs. Piper’s usage which is 0.5%)
Similar issue here and also using around 20-30% cpu (on an rpi 4) - whereas it used to use almost none.
I think you can guess where the update was… (The drops are where the rpi rebooted.)
Seeing something similar here too. On my rpi 4, it was enough to put it in a reboot loop every few minutes until I could disable the add-on. You can see the two periods where it was active yesterday and this morning:
You can use a smaller Language Model.
The default model has changed. But this is not reflected in the settings. Specify to use faster-whisper and select the model you had before.
Several libraries for working with different types of models have been merged into a single add-on. So even the project name no longer matches its essence.
Ran into the same problem today after the latest update. Whisper add-on was going into a reboot loop and pinning the CPUs assigned to it on my VM.
The fix was trivial – just upped the VM memory to 4 GB – but it was difficult to determine what the problem was. I had to manually look at resource usage for all add-ons and then check the logs for that add-on to see the OOM; it did not show up in the global HA logs.
EDIT: Actually, it did not show up obviously in the logs for Whisper either (just that the process kept dying). It was only the HA console output when looking at the VM that showed the OOM errors.
This worked for me, thanks for chiming in. I wonder if an issue should be raised in Github?
Developers just need to add descriptions to the settings page so users can better understand what’s going on.
Right now, the settings look like you’re still using whisper in auto mode.
Parakeet v2/v3 (via sherpa) are good models, but they’re over 600 MB, which can cause problems on older hardware.
Same, same. My Intel NUC didn’t know what’d hit it. Got it’s knickers right in a twist. It started interpreting Zigbee events all wrong (I’m guessing dure to the memory usage) and then it was starting the corresponding automations. My Amazon Echo kept saying things like “back door open” “washing’s finished” “garage door open” and all the lights kept turning off and on. Eventually the CPU and Memory usage was so high (from all the automations being triggered) that it caved in and rebooted.
I’ve managed to watch the NUC coming back up and restore Whisper from a backup 3 days ago before the nonsense happened again and the issues completely gone now. I guess I’ll be skipping this version or doing away with entirely.
Same here: The Whisper Add-on uses 21% processor and more memory and does not work at all.
Home Assistant Whisper logs show a repeated, critical error: “RuntimeError Failed to load model because protobuf parsing failed.” This error appears every time the add-on attempts to start, and the process exits immediately after, causing Whisper to continuously crash. This constant cycle is likely what’s creating extreme system load, as processes spawn and crash in rapid succession
Any suggestions? What can be done?
