Link to the github repo: <a href="https://github.com/picovoice/porcupine" rel="nofollow">https://github.com/picovoice/porcupine</a><p>The keyword generating application: <a href="https://github.com/Picovoice/Porcupine/tree/master/tools/optimizer" rel="nofollow">https://github.com/Picovoice/Porcupine/tree/master/tools/opt...</a><p>The keyword generating application is free for personal use only and requires a license for commercial use, but there's no pricing available and it only provides an email address for contacting about purchasing a license.<p>If anyone from the team is reading this, the information above should be front and center on the landing page. I would guess that 99% of your site visitors are going to bounce on your landing page because the relevant information is buried so deeply.
> Picovoice is a team of applied scientists and engineers who strive to build a future where our lives are enhanced with ambient voice AIs, while respecting your privacy.<p>Oh wow, thank you!<p>I also saw a reference that this is all open-source but <a href="https://github.com/Picovoice/" rel="nofollow">https://github.com/Picovoice/</a> does not have a Picovoice repository. Is it <a href="https://github.com/Picovoice/Porcupine" rel="nofollow">https://github.com/Picovoice/Porcupine</a>?
I wonder how this compares to snips[1]. I recently connected snips to my lights and it works pretty flawlessly. Picovoice does look like its easier to integrate into an app though.<p>[1] <a href="http://snips.ai/" rel="nofollow">http://snips.ai/</a>
The demo is pretty cool - the thing I liked best was this line:<p>"You can turn off your internet connection and it will keep working."<p>That's nice!
Just took a closer look at the examples and played around with some code. It works really well and a surprisingly low footprint. Basically doing exactly what is promised.<p>However i wonder why there is no way (visible?) to generate words for Javascript? Or at least a documentation on how the format for those byte arrays is build.<p>Assuming this is a licensing thing, i would really suggest to not put limit in that way. On first impression i assumed that this is useable for free for everything except commerical projects.
Is this not just <a href="https://github.com/ARM-software/ML-KWS-for-MCU" rel="nofollow">https://github.com/ARM-software/ML-KWS-for-MCU</a> trained on new keywords? Maybe I'm missing what is so special here?
How about numbers or complex/unknown words? I would love to proxy everything that isn't known AFTER the offline trigger word to Google voice recognition or something. Is that possible?
This seems really neat, but my main, blocking issue is that there is absolutely no way to add pronunciation for a wake word. Therefore, if the developers have not explicitly added the word to their vocabulary, you are completely out of luck.
It would be nice to support a grammar, a la CMUSphinx: <a href="https://cmusphinx.github.io/wiki/tutoriallm/#grammars" rel="nofollow">https://cmusphinx.github.io/wiki/tutoriallm/#grammars</a>
Which would you vote for picovoice or snips as a voice assistant AI product? I've been meaning to do more research and maybe your comments can help gain more insights.. thanks