27 comments

[ 0.20 ms ] story [ 86.8 ms ] thread
Repo has Apache license, but is loaded with binary blobs and requires paid access keys.
Title is very confusing. And there is no benchmark whatsoever to back up those claims.
Well there's this: https://raw.githubusercontent.com/Picovoice/speech-to-text-b...

So allegedly twice as accurate as Google Assistant, which sounds very impressive. No clue if it would run in real time on something like a Jetson Nano though.

Clarification. Google STT has to services. standard and enhanced. we are just better than enhanced and much better than standard. enhanced is much pricier than standard if you wonder what the diff is.

It runs real-time on NVIDIA Jetson Nano and RPI 3/4.

If you think we should consider other embedded platforms we love to hear what and why

> and RPI 3/4

Woah, now that is cool. Are you guys considering a Mycroft integration at some point?

mycroft already integrated our wake word (porcupine). we remain neutral and anyone can use our tech :)
What are the state of the art open solutions to local voice recognition? Preferably with available models that a small org can also train themselves without millions in hardware.
Title should be adjusted to:

Picovoice/leopard: On-device speech-to-text engine powered by deep learning

Since the claim is only in the HN title and not in the actually page that is linked.

Some benchmarks are here: https://github.com/Picovoice/speech-to-text-benchmark#result...

That is however not the linked page indeed.

Thank you for the clarification. `[1]` and `[2]` hyperlinks in the intro section point the benchmark
The Mozilla DeepSpeech tests on LibreSpeech listed in OPs link were out of date back in 2020[1], and Coqui.ai (the continuation of Mozilla DeepSpeech) isn't even benchmarked.

Additionally, where is the comparison to Vosk and the other noteworthy platforms? How old is the data for IBM, Azure, Amazon and Google?

https://github.com/Picovoice/speech-to-text-benchmark/issues...

(comment deleted)
Is this OSS? It looks like it has licenses and account limits baked in. Or is the release under Apache a green light to strip all that out?
Can't confirm but based on their site, I suspect that the code is OSS, but the model isn't. And that license to get the model from them is what you're getting. Not sure how they enforce or deal with the number of hours transcribed though because they say they don't collect data and the whole thing happens on device.
It would be great to have industry-recognized, concise terms for this kind of situation ("project is (FL)OSS, but contains proprietary machine learning").

For deployment of services we have "self-hosted", "cloud" and "on-prem", for example.

For ML-based projects could we have something like "teachable" (as in you can "raise" the model the way you would like to) vs "pre-directed"?

(it's been a long day; these probably aren't the neatest suggestions)

Definately not. The repo is full of precompiled dlls.
Is there anything as accurate and as small but Open Source and no subscriptions? Or at least just a small tflite model with similar specs?
it says its on device, but then mention its $899/month
> You would need internet connectivity to validate your AccessKey with Picovoice license servers even though the voice recognition is running 100% offline.

Plus closed source binaries checked into GitHub which is always a red flag; you're basically using it for free distribution of your paid software.

Paid software on github, this post is just an ad.
steps to try it out:

- sign up for an account to get an access key - install demos with `npm install -g @picovoice/leopard-node-demo` - run `leopard-mic-demo -a $ACCESS_KEY`

I'm pretty impressed with the accuracy, especially given it's all local and a 20MB model (though the total weight with node modules included comes out to 36MB). Obviously the licensing restrictions are a bummer, but 100 hours a month should be more than enough for my purposes - I like to compose thoughts with my voice sometimes and capture something as close to the stream of consciousness as possible. Really curious to see what gets built on this.

Accuracy is important but latency plays an important role for viable speech recognition.