[dsteinman]

I am attempting to bring 100% client-side speech recognition to the web:

https://github.com/jaxcore/bumblebee

Although the first release is not officially out yet, the NodeJS code is working and you can install the development version of the app server and try out the hello world app locally.

The solution involves running Mozilla DeepSpeech inside an Electron desktop application with a websocket server and client API that NodeJS scripts can interact with, to receive speech recognition results, utilize "alexa" style hotword commands, and text-to-speech. The electron app handles all the heavy stuff, and you just use a simple API.

A web browser extension can also make use of this API to bring these capabilities to web sites, but that part isn't finished yet.

[IshKebab]

It's not really "the web" if you have to use Electron and Node surely? Wouldn't it make more sense to do it with web workers and wasm?

[dsteinman]

The web browser extension would communicate with the electron app server, NodeJS would not be needed in that scenario (the electron app includes the nodejs server code). You can write your web voice app with static client-side JavaScript which communicates with the Electron server through the browser extension.

Web Page <-> Bumblebee JS API <-> Bumblebee Extension <-> Bumblebee Electron App (DeepSpeech)

DeepSpeech with the pretrained english model is enormous (1.4GB) it's not feasible to load it into a web worker. It can run in a server, but then every website would have to run its own server side speech recognition servers which is difficult and expensive to scale.