[BUG] sil causes word-pronunciation mismatch only when training

[jeffmielke]

Debugging checklist

[x] Have you updated to latest MFA version?
[x] Have you tried rerunning the command with the --clean flag?

Describe the issue
A clear and concise description of what the bug is.

Training acoustic models ends in an error like KeyError: ('a', 'G UH1 D') where the p in pron_mapping[(w, p)] is always for a word that appears after the w in the corpus ("a good" in this case). This seems similar to #471, because the problem goes away if I remove all sil pronunciations from the dictionary. But this issue only seems to happen when training.

For Reproducing your issue
Please fill out the following:

1. Corpus structure
   • What language is the corpus in? English
   • How many files/speakers? 30/42
   • Are you using lab files or TextGrid files for input? TextGrid
2. Dictionary
   • Are you using a dictionary from MFA? If so, which one? no
   • If it's a custom dictionary, what is the phoneset? arpabet
3. Acoustic model
   • If you're using an acoustic model, is it one download through MFA? If so, which one? n/a
   • If it's a model you've trained, what data was it trained on? trying to train on a subset of the Raleigh corpus

Log file
Please attach the log file for the run that encountered an error (by default these will be stored in ~/Documents/MFA).
train_acoustic_model.log

Desktop (please complete the following information):

• OS: [e.g. Windows, OSX, Linux] Linux
• Version [e.g. MacOSX 10.15, Ubuntu 20.04, Windows 10, etc] Ubuntu 20.04
• Any other details about the setup (Cloud, Docker, etc)

Additional context
Add any other context about the problem here.