n-grams

This program uses diphone (bigram) and triphone (trigram) models to represent Mainstream American English (MAE) phonotactics.

Output:

1. Generate random sequences of phonemes based on training data (CMU Pronouncing Dictionary)
2. When given test file of made-up "words" as argument, model scores each "word" based on similarity to MAE phonotactics

Use:

The program takes either two or three command line arguments:

1. training_file

   • CMU phonetic dictionary

2. N (either 2 or 3)

   • sets n-gram type (bigram or trigram)

3. test_file (optional)

   • corpus for which to calculate perplexity

     • X.txt or Y.txt

Notes:

If the program is given only two command line arguments, it prints 25 "words" consisting of random phoneme sequences based on either a diphone or triphone model.

When given three arguments, the program processes the test file based on a smoothed di- or triphone model. The probability of each test "word" is calculated and printed. The perplexity of the test corpus is calculated based on the log probabilities of its constituent words.

ngram-TTS.py will only take two arguments, but it reads aloud each randomly generated phoneme sequence / word using the pyttsx3 library.

Known Issues:

• triphone counts not accurate due to misapplied boundary symbols (# #)
• N in perplexity calculation is defined as len(phonemes), but it should be len(phonemes)-1