Yes. The big parts are the models (neural networks (NN)). You can make them quite small (sparsification, quantization, transfer learning to smaller models, ...)

Yes. The big parts are the models (neural networks (NN)). You can make them quite small (sparsification, quantization, transfer learning to smaller models, ...). But this is non-trivial. Kaldi is the most complete open source speech recognition software, covering HMM/GMMs, the conventional hybrid HMM/NN, and much more. If you care about end-to-end only, there is much more, e.g. Mozilla DeepSpeech, and most modern open source encoder-decoder-attention frameworks can do that (e.g. Lingvo etc). Most of these encoder-decoder-attention framework can also do text-to-speech (TTS). Maybe you need a separate vocoder, but you will also find open source code for that. For the assistant engine, there are also a couple of open source projects, e.g. Mycroft, LinTO, Rasa, Leon, Open Assistant, Susi, ... Putting it all together is non-trivial, and requires lots of expert knowledge, and lots of work and time to get a good result. This is hard to do as a hobby project.