I’ve been reading this post and added a Pull Request for them to add Alexa wake word to the Voice PE official configuration.
The only thing missing is that I don’t know how to test False Accepts per Hour values against all units and channels from the Dinner Party Corpus, so we can add real values to the sensitive selectors. I am investigating in the meantime how to do that.
Hopefully this can be configured correctly and be merged soon. I’ll try to help the best I can to make this happen, since I tested the changes in my Pull Request and they seem to work correctly on my Voice PE I have at home.
I don’t see the ever adding some elses (probably copyrighted) wake word to thier stack… Ever. It’s a lawsuit waiting to happen. In the US you defend your copyright or loose it, therefore Amazon would be forced to file.Would be almost immediate. Same story with Cortana or Google.
Oh I didn’t think about that, but yeah they could be liable to a lawsuit I see that Alexa is not copyrighted, but it’s trademarked and they could be sued because of “trademark infringement”
Yup copyright, trademark, service mark, anything remotely associated with the name of… Yeah immediately. Within 24 hrs most likely. Especially with them trying to finally monetize Ale… She who shall not be named…
I have a TTS and STT working, the issue is the wake word. Ironically, I am trying to make a GLaDOS one as well. I added this to my config and it keeps getting stuck at this step:
[15:17:55][D][esp32.preferences:142]: Writing 1 items: 0 cached, 1 written, 0 failed
Would you be able to provide your full YAML (without the key of course) so I can make sure everything is right on my end? Also, was there anything specific you had to do before flashing it? Any help would be greatly appreciated.
This is actually a really interesting question. The “Alexa” trademark is for goods and services. ALEXA Trademark of AMAZON TECHNOLOGIES, INC. - Registration Number 6511645 - Serial Number 86558550 :: Justia Trademarks A wake word is neither. If a company branded its hardware or software as “Alexa” or something similar, then yes, definitely infringement. But neither HA’s product (“Home Assistant Voice PE”) nor any HA-approved voice services (“Whisper” “Wyoming Protocol” “Piper”) is so named. Moreover, aren’t they open source?
Amazon offers a few optional wake word choices, such as “Ziggy” and “computer”. Neither are trademarked (nor could “computer” be trademarked), as neither are used to identify a good or service.
The real question is whether there would somehow be a likelihood of confusion for the public if HA offered its users the option to use “Alexa” as a wake word. I suppose Amazon could argue that it would tend to make people think that HA is affiliated with or licensed by Amazon, but if the default wake word is not “Alexa” and it requires the user to make the change, then is it HA or the user that causes the confusion? Plus HA could have a notification pop up and/or verbalized if a user selected that option, something like “HA is not affiliated with Amazon and your choice to change the wake word to “alexa” will not cause your HA system to be connected to Amazon’s Alexa service. Yada yada yada”
Of course, nothing prevents anyone from filing a lawsuit and I hear Amazon has a few sheckles saved for rainy days or flimsy lawsuits.
Its not hard. 100% on the user. If they don’t provide it they’re off the hook what you do with your system is your business.
If Nabu did it- lawsuit immediately. Because in the US if you do not defend a trademark you loose it.
If a private citizen did it… As long as they aren’t providing it to others for a fee. It’s debatable of there’s even a case…
(source I used to be an escalation manager for one of the big 3 tech shops, a variant of this question was in my training pack. Business cant do it theyll get lawyerd up immediately.)
It’s really easy to do it yourself, and it’s “unofficially” included (see examples above in this thread).
If they shipped it as part of a release, and documented it, Big Rainforest Co™ would come after them, not to “win”, but to defend their trademark name (even if it is defined as a “good/service”).
I think that Nabu Casa should come up with a new, unique wake word name, train a really good OWW model on it, and then ship it with the next revision of their Voice hardware. “Okay nabu” doesn’t exactly roll off the tongue. (And, similarly, I’m not a fan of “Okay google” either.)
I was successful generating a tflite model with this, but had to install wget and ffmpeg@7 via homebrew (this is on a new M4 pro macbook running tahoe and I don’t have much of a dev environment setup yet), and I needed to add export DYLD_FALLBACK_LIBRARY_PATH=/opt/homebrew/opt/ffmpeg@7/lib/
just above the following line in train_microwakeword.sh (around line 220) in order to get libtorchcodec to recognize the correct location of libavutil.59.dylib: python scripts_macos/make_features.py
Haven’t tested it yet in a esphome config, that will be a project for another day
Agree 100%. Provide a selection of names no longer in use (Michelangelo would be awesome). Go biblical (Metatron is my suggestion).
Ooooooor - just make wake word modeling part of the device itself. Make it easy for the end user. This would avoid and any all Trademark issues altogether.
Agree. Alexa official is probably not the best idea. Alexa as a hobby project for each home user is the way to go. We just need a much better wake word trainer in HA. Hardware is an issue as I would presume most people don’t run HA on a machine with a top-tier GPU, but if we could be more confident that a solution works, I wouldn’t mind as much if the trainer ticked away in the background for a day or two. My issues were frustrating because it took hours even on my machine with an NVIDIA 4060, only find it didn’t work and I had to re-train over and over.
There is ZERO way around training without a GPU. 4060is fine IF you have the ram for the model. You must have the gear my issue is even with the gear the software stack is still prohibitively hard for the average human. It’s doable. But. Yeah.
What about an add-on that guides a user on various options for training that are supportable? For instance we could verify a few different options that actually work the first time, with offical instructions in the add-on? For instance sections for well-tested methods and links to those methods and pro/con for users. Include sections for docker mww trainer in raw python and Jupyter, with explainations for GPU requirements and time estimates. Maybe an AWS CloudFormation template or SageMaker instructions, and whatever the Azure equivalent is. Track it in a repo like anything else.
I don’t mean to send this discussion sideways and I’m not at all familiar with US trademark/copyright law, but Cortana is just a name, as is ‘she who must not be named’ and you can’t trademark or copyright a name.
As an example, in the sci-fi tv show HALO, they had an AI character called Cortana. Would that not be classed as a breach if the name and use of that name was trademarked or copyright?
Cortana as a character is a wholly owned property of bungee > whatever studio is holding Halo for MS> Microsoft. Microsoft paid ms Jen Taylor a ton of money for her work.
You can Call something Cortana but make it blue and glowy - MS would have an issue just like Disney does with Mickey Mouse.
I see that Alexa is not copyrighted, but it’s trademarked and they could be sued because of “trademark infringement”
Apple Silicon Version: 
