Mozilla Deepspeech?

[ghost]

I think this is potentially some big, big news for OpenSTT and Mycroft: https://github.com/mozilla/DeepSpeech

My reason for not using Mycroft, and I think many people's, is refusing to send my speech to Google. If DeepSpeech could offer a local or self-hostable alternative, even if it required set-up on another machine with a GPU etcetera, then I'd jump at the chance to try Mycroft.

[clusterfudge]

So are you specifically opposed to sending data to Google, or all remote services? I believe the plan is for Mycroft to host their own, and open source the stack/trained models. Even given that, the requirements for running such a service may be outside the capabilities of a hobbyist (a large portion of Mycroft's following).

[ryanleesipes]

Does DeepSpeech also have the training data they are using included? Will they release the models created via their efforts? Looking through the repo now, trying to find this info.

[ghost]

They have said in an issue on the repo, that they will release trained models soon. In answer to an earlier query on this subject: I don't believe *any* organisation has the right to listen to everything I say, or to know how I interact with my devices. Google are obviously factors higher on my scale, because their entire business model is the exploitation of data obtained by spying on me. But, even if there were a Mycroft.ai-run service, I'd still be leery of it, and desire my own. I'm aware that most people don't really care about their right to privacy, until they suffer for it. But for those of us that do care, the ability to self-host is a critical feature.

…

-- @onetruecathal / @cathal@quitter.is

On Thu, Mar 23, 2017 at 7:52 PM, Ryan Sipes ***@***.***> wrote: Does DeepSpeech also have the training data they are using included? Will they release the models created via their efforts? Looking through the repo now, trying to find this info. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[clusterfudge]

@cathalgarvey While I don't disagree with your sentiment, I think you may be conflating surveillance vs. a voluntary exchange of data for services (or contribution to a community). I would prefer to avoid a heated discussion on how various organizations may abuse that difference, but I would like to call out a point in favor of the latter with respect to Mycroft.

Mycroft (and Mozilla) are open source organizations whose goal is to provide back both software, and in the case of machine learning products, trained models, to the community. As a member of the Mycroft community, you could choose to provide your voice data. This in turn would allow Mycroft to construct better models and train them more effectively. The development of a model has the prerequisite of a substantial amount of training data, so I'm sure your help would be greatly appreciated in that respect.

Back to the nuts and bolts of Mycroft :)

Mycroft supports multiple STT providers out of the box, and if you can get an instance of Mozilla Deepspeech running, it should be pretty trivial to add your provider to the mix.

See here for Mycroft's STT client code.

[sixtyfive]

How about an opt-out option, which would make it possible to download a snapshot of all training data to your self-hosted service? That way nobody would have to set up some complicated service, but everybody who wants to do so, could.

[ryanleesipes]

From what I understand they are releasing open source models that can be run locally.

[ghost]

Whenever they do, it'll be great to get it in Mycroft.

Just to address your points @clusterfudge - My objection isn't with a "voluntary" sharing of voice data with an organisation I ~trust, such as Mycroft. Although, in practice, I would choose other ways to support the project. My objection is that at present, I can only use Mycroft if I also share my environment with organisations I explicitly distrust, such as Google or Microsoft, etcetera.

So, while Mycroft is an exciting project that I'm watching carefully, I cannot use it without subjecting myself to surveillance. And that means I won't use it, at all. If a module for a non-surveilling STT system becomes available, I'll be installing/using Mycroft immediately.

[nealmcb]

Watch mozilla/DeepSpeech#435 for updates on pre-trained models for DeepSpeech.

[ghost]

Based on notes in the Deepspeech repo, they're planning a first release this month! And, it appears there may be beta-models in the repo (fetch with Git-LFS) already. So this might be ready to start building with.

Where does one begin, making a STT back-end for Mycroft? If I wanted to glue Deepspeech to Mycroft, where should I start?

[nealmcb]

And indeed, their first pretrained model is released!
https://github.com/mozilla/DeepSpeech/releases/tag/v0.1.0
https://twitter.com/nealmcb/status/934428278722510848
It is supported on Raspbian Jessie on Raspberry Pi 3.

Now I guess we should be documenting how to set up a Mycroft server that uses it, making it easy for folks to configure their systems to use it, and training it on the Mycroft Open Data Set!

[nealmcb]

I keep hearing that Mycroft is engaged with the Mozilla folks on DeepSpeech, but I'm surprised to not see more news on that front, or requests for help, or the like.

See mozilla/DeepSpeech#847 for info on the deepspeech-server code, and movement towards integration.
What else is going on?

[jgbreezer]

Having read the github deepspeech-project README, it sounds like you might have to choose a trade-off of a longer delay (maybe quite notably?) between speaking and it knowing what you're saying unless you have considerable money (maybe thousands of pounds sterling?) to spend on hardware (or on-demand GPU cloud services to apply the machine learning).

[ghost]

Privacy isn't free :) -- Sent from my Android device with K-9 Mail. Please excuse my brevity.

[nealmcb]

@jgbreezer, At the moment, the DeepSpeech models that Mozilla is using are much less mature than the ones used by the big commercial services. I think we can expect that to change pretty quickly, and be combined with the burgeoning market for low cost deep neural net hardware for use on mobile devices as well as desktops, resulting in fast, private, inexpensive, good quality speech recognition for all.

In the meantime, in order to get there sooner, we can help out by choosing to share voice samples with the open data communities, as Mozilla invites via their programs, and as Mycroft is doing via the "opt-in" program. See more at Privacy and Machine Learning | Our Open Data Set and Opt-In Feature - Mycroft.