Picovoice – Embed private voice AI into any product instantly

https://picovoice.ai/#voice-control-demo

182 75 8 years ago

anc84• 8 years ago

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.

Oh wow, thank you!

I also saw a reference that this is all open-source but https://github.com/Picovoice/ does not have a Picovoice repository. Is it https://github.com/Picovoice/Porcupine?

xf86alsa• 8 years ago

Porcupine is just wake word detection, which in itself is very useful indeed.

Not sure what the plans for open sourcing the rest of the components are though.

kenarsa27• 8 years ago

Hello. Thanks for the comment. We have plans to open source two more products this year. with similar licensing compared to Porcupine.

1- Speech-to-Intent: It allows you to issue complex voice commands in a specific domain and in turn returns the intent. For example, in the case of a coffee maker you can say "Please may I have a single shot espresso with no milk and two sugars". The engine returns a JSON-like object with {"product": "espresso", "milk": "no", "sugar", "two", "# shots": "2"}. It is a tightly coupled domain-specific speech recognition and NLU. It is small (less than 3MB and 8% CPU usage on RPi3) and ideal for home automation, industrial application, service industry, etc.

2- Speech-to-Text: It is large vocabulary speech recognition software that runs locally. It will support all platforms currently being supported. It allows you to do large vocabulary transcription with high accuracy locally on an embedded platform.

tyingq• 8 years ago

The open source reference is this phrase: "Picovoice's speech to text repository ranks among top 10 open source machine learning projects."

It's apparently only talking about their benchmarking tool: https://github.com/Picovoice/stt-benchmark

The engine itself is called Cheetah, and appears to be a closed source, "inquire for pricing/licensing" product.

kenarsa27• 8 years ago

We will open source Cheetah this year with a similar licensing model as Porcupine.

kenarsa27• 8 years ago

Stay tuned. We will add two more repositories (1) speech-to-intent (2) speech-to-text

pault• 8 years ago

Link to the github repo: https://github.com/picovoice/porcupine

The keyword generating application: https://github.com/Picovoice/Porcupine/tree/master/tools/opt...

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.

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.

kenarsa27• 8 years ago

This is Alireza. I am the founder of Picovoice. Thanks a lot for the information and also sharing. I 100% agree we do need to work on the website and its content. It is definitely in our TODO list. Will prioritize. In the meantime happy to answer any commercial questions via email contact@picovoice.ai

Thanks again for sharing

edoceo2• 8 years ago

You could have answered the pricing question here, directly, in response to someone asking about it.

You said it needed to be addressed, and was a priority yet the question remains unanswered. It gives me the cognitive-dissonance.

toniprada• 8 years ago

This is perfect for a Home Assistant (hass.io) integration.

roel_v• 8 years ago

Can I use this for my home automation setup? It seems the github link is only part of the product. Do you license this for individual users?

ocdtrekkie2• 8 years ago

This is my question too. I looked at Snips but they had no timeline on Windows support and were more interested in talking about their crypto coin than answering basic questions about their software.

azadal• 8 years ago

Where were you enquiring on? I've had a lot of help from their discord channel. You could give that a try.

syntaxing4• 8 years ago

I had pretty good luck reaching out to the developers on Discord these past couple weeks. Not sure when you tried but I recommend trying it again. I personally I think snips is way more appropriate on something embedded like a RPi. I was up and running with their new sam package within minutes. Their new update (about two months?) really made it more user friendly. Unfortunately, their windows packages is still very broken.

E: You should probably try in the late night or early morning time for the States (EST) since they are located in Europe I believe.

kenarsa27• 8 years ago

We do support Windows!

kenarsa27• 8 years ago

You can use this for free under Apache 2 license for a personal non-commercial use. No licensing needed. You don't need to pay us a penny as long as you don't make money off it.

gpm2• 8 years ago

The Apache 2 license allows for non-personal and commercial use...

gpm2• 8 years ago

Is this related to picotts?

syntaxing4• 8 years ago

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.

[1] http://snips.ai/

kenarsa27• 8 years ago

I am the founder of Picovoice. I am NOT going to comment on comparison as obviously I will be biased ;)

With Picovoice you can use the voice control engine to accomplish this. Maybe something similar to this demo?

https://picovoice.ai/#voice-control-demo

The cool thing about this engine is that it is tiny. It uses less than 8% CPU on RPi3 and altogether it is less than 2MB (code, model, etc). Technically you can run it on something much smaller and cheaper than RPi.

Alternatively, the speech-to-intent engine could be a good candidate. More information on this along with an interactive demo will be released this weekend.

MrBuddyCasino• 8 years ago

How much work do you think will an ESP32 port take?

syntaxing4• 8 years ago

I would love to see some benchmark numbers on Picovoice. The small size and low CPU is definitely interesting but I'm worried that performance is hindered because of this. Also, is 8% peak usage? Having this loaded on some small IoT chip will be awesome to see.

kingosticks5• 8 years ago

Your website says:

Runs in real-time with only 5.6 MB of memory and 25% CPU usage on a Raspberry Pi 3.

frontier87• 8 years ago

I think whichever privacy centric voice assistant will come to market first will build an unstoppable first mover advantage. There are so many people just waiting for an alternative to the Googles/Amazons of the world.

Snips looks like a promising prospect!

sctb• 8 years ago

So you said in your only other comment. Hacker News is a community site and it's not just to promote something, so could you please participate community-style instead?

https://news.ycombinator.com/newswelcome.html

MR4D2• 8 years ago

The demo is pretty cool - the thing I liked best was this line:

"You can turn off your internet connection and it will keep working."

That's nice!

kenarsa27• 8 years ago

It is true! Have you tried it? :)

We will add another demo for speech-to-intent module fairly soon (over this weekend). Stay tuned!

MR4D2• 8 years ago

Not yet - my work computer doesn't like being cut off from the world due to financial apps that I run, but I'll try when I'm home this weekend.

KaranRaut• 8 years ago

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

natvert• 8 years ago

Is this not just https://github.com/ARM-software/ML-KWS-for-MCU trained on new keywords? Maybe I'm missing what is so special here?

kenarsa27• 8 years ago

the project you mentioned is wake-word for ARM only. Great project, BTW. For wake word, we provide on-demand model generation. We do also more than wake word. Finally, we can run on other CPUs/OSs as well.

bmc75052• 8 years ago

It would be nice to support a grammar, a la CMUSphinx: https://cmusphinx.github.io/wiki/tutoriallm/#grammars

kenarsa27• 8 years ago

Agreed. The reason for not supporting grammar in this product is that we want to keep it extremely lightweight. We do have upcoming products that support grammar as well.

bmc75052• 8 years ago

Awesome! I look forward to trying out the speech-to-intent and speech-to-text products. Will these also be Apache 2.0 licensed?

bufferoverflow• 8 years ago

Doesn't work on Android chrome.

kenarsa27• 8 years ago

The demo is using WebAssembly which is supported by Chrome. It also uses Web Audio API which I believe is again supported by Chrome. I just used the demo on my Android phone using Chrome. I wonder what could be a problem. It the mic on? :) If yes, maybe provide me the version of chrome you are using. I will look into it.

herbst4• 8 years ago

It works. But didn't on the first or second try for me. Suddenly it just worked. Dunno

lucb1e2• 8 years ago

Works in Fennec (Firefox but from f-droid), but it took a sec to load the library it seems. Maybe that's the issue?

jscheel• 8 years ago

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.

kenarsa27• 8 years ago

If you have a commercial application we can build you the model for any wake word. Since this requires some engineering work on our side we can only offer it to commercial customers at this point.

lucb1e2• 8 years ago

Why not make it possible for anyone to train it instead of you having to do engineering? Is that the profit model? If so, totally valid, but I'm wondering if I'm understanding this right.

herbst4• 8 years ago

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?

kenarsa27• 8 years ago

Numbers and complex words (I am assuming you mean something like "ok blah"?) are doable easily. I am not sure what you mean by unknown words. Could you elaborate?

Obviously, you can just grab the audio stream after the phrase is detected and route it to whatever you like. Google ASR or even a local one running on the device!

herbst4• 8 years ago

That is exactly what I ment if I get the audio stream out as well.

I have literally 3 projects, one a current big that depend or would heavily benefit from this. Can't wait to give it a try!

arendtio• 8 years ago

@kenarsa any plans for a Go integration?

Pretty impressive tech by the way.

herbst4• 8 years ago

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.

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.

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.

simonvc• 8 years ago

Came here to say this.. Had a great idea for using this to help train pilots to do the RT (i'm doing my PPL now, and this would be perfect..) The lack of JS optimizer means i can't really hack on what, could have become a nice passive income site.

Picovoice – Embed private voice AI into any product instantly

Related Stories