Vosk 0.3.30, SetWords support, acceptWaveform with "raw" float buffer

[Yahweasel]

Hello,

I made some updates to vosk-browser that I needed for my own (potential) use of it. Hopefully none of them are too offensive, so you may want to integrate them.

1. I updated Vosk to 0.3.30. This was more painful than it should've been. 0.3.30 really wants a newer version of OpenFST, but trying to update OpenFST caused Kaldi not to compile, and trying to update Kaldi caused the MKL dependency to break, and I couldn't get the newer MKL to compile under Emscripten. Luckily, then OpenFST issue is incredibly trivial, so instead I just included a one-line patch to vosk, with which 0.3.30 works fine.
2. I added SetWords. This involved adding a class of "set" messages, which I generalized so that newer recognizer settings can be added in the future, though the only one I added was words. The words setting is preserved in the frontend "meta-recognizer" (whatever you'd like to call it), so that if a new backend recognizer needs to be made because sample rates change, it will preserve the setting.
3. I added a version of acceptWaveform that works with a raw float buffer, acceptWaveformFloat. I need this because in my project, I already have three or four other components hanging off of an AudioWorkletNode, so conforming to AudioBuffer or your AudioWorkletNode was not an option, but just giving a waveform was.

If you have a better solution for upgrading Vosk, it would obviously be preferable :). Everything else here is pretty tame.

[nshmyrev]

0.3.30 really wants a newer version of OpenFST, but trying to update OpenFST caused Kaldi not to compile

Not just newer, we have our own version with important speed fixes (about 20% faster). It is better to build our openfst than generic one.

[Yahweasel]

0.3.30 really wants a newer version of OpenFST, but trying to update OpenFST caused Kaldi not to compile

Not just newer, we have our own version with important speed fixes (about 20% faster). It is better to build our openfst than generic one.

OHHHHHHH. I'll bet that's why that was such a rabbit hole :). My apologies, I was not familiar with that quirk.

[ccoreilly]

Thank you very much for your contribution @Yahweasel ! I have been meaning to update vosk-api to the latest version (as well as adding other stuff) but have had no time, so it is really appreciated.

As @nshmyrev points out, vosk-api uses its own fork of Kaldi and OpenFST but when I started working on vosk-browser I had issues with OpenFST 1.8.0 compilation as it would not find the zlib dependency. I'll try to fix that again.

I'll test your changes locally and merge/publish them, thanks again!

[ccoreilly]

FYI: I updated the build and it now uses the OpenFST fork of Alpacephei in version 1.8.0. I will publish a new package in npm once I've cleaned up some things.