Hacker News Reader

Comment by tmzt | Hacker News Reader

How does this compare with Snips.co which can do offline speech recognition on a Rapberry Pi 3?

Coyld this be used to train a model/engine that can be used that way?

The model we released today is not yet optimized for smaller devices like that, but our plan is to make it usable on targets like the RPi3.

Are you releasing any prebuilt models, I searched but couldn't find any, so people can go and play with your work without training?

Edit: NM found it under releases: https://github.com/mozilla/DeepSpeech/releases.

Thank you for finding and linking the prebuilt model. It was eluding me.

OMG, C++??? I thought Mozilla folks would give Rust a spin.

It is Tensorflow based which has a C++ API. Though, it looks like they provide Rust bindings for the Deep Speech library itself.

I can't find exact numbers on Snips.ai, but generally there's a linear relation between the size of the inference model in RAM and the accuracy it can obtain.

I'd have to assume DeepSpeech outperforms anything running on a RasPi3, at least for LVCSR. It hits 93.5% accuracy on Librispeech, which I've never seen from any offline recognition models.

Kaldi has 4.14% WER (95.86% accuracy) on the same test dataset (test-clean) [1] using a model that runs faster than real time on CPU. You would have to make the model smaller to run it in real time on a RasPi3, but according to this [2], you can get decent WERs for read speech even then.

[1] https://github.com/kaldi-asr/kaldi/blob/master/egs/librispee...

[2] https://groups.google.com/d/msg/kaldi-help/Pr6jPH1Qshg/kn8df...

They are listed as a contributor to the Common Voice dataset: https://medium.com/mozilla-open-innovation/sharing-our-commo...

I think their product is higher level - it has things like recognizing commands and sentence structures, and hotword detection.

The URL is https://snips.ai/ .

Thanks, going by memory which is apparently faulty.

I don't think snips is open source, for one.

Reply on news.ycombinator.com