How many new words and phrases could be added in real time without effecting the whole recognition accuracy.

### Issue Title:
Questions about Adding Phrases in the Dynamic Graph Update

### Issue Description:
In this example, we demonstrate how to add phrases in the dynamic graph update using the following code snippet from the [Vosk API](https://github.com/alphacep/vosk-api/blob/master/python/example/test_words.py):

### Python
rec.SetGrammar('["one zero one two three oh", "four five six", "seven eight nine zero", "[unk]"]')
or:
rec = KaldiRecognizer(model, wf.getframerate(),'["one zero one two three oh", "four five six", "seven eight nine zero", "[unk]"]')
### C++
void Recognizer::UpdateGrammarFst(char const *grammar)

**My first question is:**
Do all those words ("one zero one two three oh," "four five six," "seven eight nine zero," "[unk]") need to be present in the lexicon HCL.fst (words.txt)? Should we create a phone sequence to enable the recognizer to handle these unknown words?

Is it possible to feed this list of sequence sentences or words from a file rather than hard-coding them as a list?

**My second question is:**
How many phrases can be added in real-time without significantly affecting the overall recognition accuracy (optimal number ~)?