speech recognition for dictation which works for sailfish?

answered 2020-02-03 04:28:16 +0200

DrYak
9624 ●93 ●134 ●128

http://www.sympato.ch/

updated 2020-02-03 11:38:03 +0200

Currently there are few speech recognition systems which aren't running on some neural net AI (LSTM).

Sphinx has a mobile-friendly system called PocketSphinx, that runs some simpler machine learning system that can fit in a mobbile setting and can be trained on a limited subset of the language corpus.

(But for other mobile-friendly see at the end).

The rest are all AI neural net based, which is much more ressource intense.

• most systems (including the mobile version of Dragon dictate) just stream their audio inputs to some cloud server (Google is another example, and on openrepos you can find Saera for Sailfish which also is cloud-based).

• a few of the ultra modern high range phones have built-in accelerators that can help with the kind of matrix maths used in neural nets. Saddly none of them are currently running Sailfish OS.

• the last alternative would be to run your own server. (either on computers with GPUs, or on special SBC that have an ARM with an AI accelerator - see RK1808 based AI Module such as, e.g.: pine64's SoEdge)

• a fall back instead is to fall back on mobile-friendly (see below).

There are opensource neural-net based technologies:

• Mozilla has been doing work on opensource technologies (DeepSpeech), they have introduced revently a TFLite version that can also run on ARM.
• Mycroft has also opensource codecode, that can be run on ARM for the speech-to-text part, it's partnering with Mozilla (and in the past has used pocketsphinx for the wake-word recognition).

There are also cloud provider who can propose you to spin up closed-source cloud instance dedicated for that:

• I think I've seen it proposed on Microsoft Azure (you can get your own speech cloud VM), but I haven't much experience (professionally, we're a research group using more classical statistical analysis. Little NN done currently).