Investigate a better speech-to-text module

[gcampax]

Bing Speech is good but not exceptional quality, and the streaming API is going down, leaving only the REST API which has no feedback while the user speaks and is also quite slow.

We should investigate alternative software, such as Mozilla's deepspeech. If necessary, we can run our own servers, which is probably a good idea for privacy anyway.

[12people]

Open AI just released its high-quality speech recognition model under the MIT license: https://github.com/openai/whisper .

It offers several model sizes with different system requirements, but all potentially usable offline. The ideal solution might be for Almond to ship with a smaller model (or feature a setting allowing users to choose to download an offline model) and run a server with a larger model.