Could you try again? 
It works now
For anyone reading, I tried this for the first time with the parakeet v3 multilingual models and it’s 4 o 5x faster than whisper turbo while being as accurate or (subjectively) even more accurate, at least in spanish.
It’s even usable on CPU (~1s on my intel nuc with an i3).
After a quick test as an addon I ended up installing as a docker container so I have full access to my GPU (RTX 3060) and it’s insanely fast, we’re talking 0.05s per command, while whisper-turbo took ~0.35s and, IMO, had worse results.
So far the only downside I’ve been able to detect is that it uses more vram than whisper-turbo, which is a bit surprising since parakeet it’s a 0.6B model.
2 Likes
Thanks! I switched on this recommendation. Whisper small-en was 3s and the default onnx-asr model (models–istupakov–parakeet-tdt-0.6b-v2-onnx) is now 0.42-0.59s on intel i5 4th gen cpu.
It looks like the official whisper addon has added support for parakeet: addons/whisper/CHANGELOG.md at 86e40f6b5d50ac0a2643d00b7dc8bc18879ca81e · home-assistant/addons · GitHub
Would someone be able to do a comparison and check if it performs the same? (I.e. is this addon obsolete for the key feature)
I won’t stop supporting this addon, but this addon might only be useful for trying more obscure onnx based models.
I’m using this in the docker variant since this can be gpu accelerated which is not the case for the official addon i guess?
Whats the performance of nemo-parakeet-tdt-0.6b-v2/3 in german language compared to whisper?
I tested both whisper addon implementation and your addon, and my impresion is while this addons comsumes more ram seams to work better for me at detecting what i say if i am bit further away from the mic, so i will stick with your addon thanks.
The new Canary-1B-V2 model has been released. Is there any chance of its implementation?
@mchk Do you have container version available for your addon please ?
1 Like
I’ve never created dockers before how would I go about creating one for unraid if you do mind helping me out. I think I would like to try this out I have a nvidia card on the server
Anyone know if the new Nemotron ASR which released this January will be made available by either of the STT addons ?
but I can’t check the quality of recognition.
Some people think the model isn’t very good, and the offline parakeet v2 model is better. So if OnnxASR works at an acceptable speed, then there’s no need to waste time on it.
If you test it, please share your results.
Hi, I’m noticing that my Home Assistant backup size from about 500mb has gone to 2,5gb. After searching it was a8xxxx_onnx-asr.tar data. Correct me if i’m wrong but in the documentation it says that this data will not be included in the backup? Did I miss something or do I have to make some changes?
Thanks in advance for the help.
Sorry about that, just an error on my part, should be resolved in 0.3.1 which has just released.
1 Like
Thanks a lot for the quick fix 
Hi, sorry to tell you, but even after the update, it is still backing up the llm file.
I have already restarted HA, but that didn’t help. I’m running HA on bare metal, so no other stuff.
Rebooted HA. Full backup without ONNX ASR: 394 MB
Full backup with ONNX ASR: 2.6GB
Maybe this will help?