Create a reading list

Your task is to maximize diphone coverage in a list of 500 sentences. The coverage is 100% if all diphones appear at least onece in your reading list. You can use any method/algorithm you like to achieve this as long as you motivate your design choices.

Your main goal should be to achieve a better coverage than by simply choosing a random set of 500 sentences which yields a coverage of approximately 67%.

Your 500 sentences can be sourced from ./data/list.tsv. Each line in this file has the format <sentence>\t<pronounciation> where:

• <sentence> is some piece of text
• <pronounciation> is the corresponding phonetical prediction. Note:
  • Each <pronounciation> ends and starts with a ~ symbol to denote start and end of sentence.
  • Each <pronouncation> is split with a \t symbol between words.

The basic outline of your algorithm will likely look something like:

1. Create a list of types that you want to cover (for example all diphones)
2. Find a sentence in the corpus which provides the largest number of different types in the wishlist.
3. Add it to the script
4. Remove the types from the wishlist
5. If the script is 500 sentences, stop. Otherwise go to step 2.

Most of the work will go into step two where you have to create some sort of a cost metric for comparing sentences from the corpus. Ideally your cost metric will also consider the length of the sentences.

Adding your own sentences

You might want to add your own sentences to the list. by adding e.g. your name to the list of sentences you would guarantee that your TTS will always output your name in the correct form. But to do this, you have to first run a G2P model on your custom sentences. I have created a repository for you to do this and it is available here.

What to turn in

• Turn in your code as usual
• Turn in the reading list you generated by using the tools.save_list() method. See example.py for usage. Furthermore, aach line in the list should have the form <sentence>\t<source>\t<score>\t<pron> where:
  • <sentence> is the sentence being read
  • <source> should be your RU id (e.g atlithors for the email atlithors@ru.is)
  • <score> a score that dictates in what order your sentences will be prompted. A higher score should mean that the sentence appears earlier. So for 500 sentences you could score them with e.g. 500, then 499 etc.
  • <pron> The pronounciation prediction in the same format as used in ./data/list.tsv
• A document explaining your design decisions:
  • What is the unit type you are trying to cover? (e.g. diphones)
  • How your algorithm works.
• clearly state the coverage your script achieves.

Notes:

• Picking the longest sentences is likely to be a good strategy (since they will contain the highest number of phonemes) but do note that you will be reading the sentences. Take some time to carefully consider how you could maximize diphone coverage while minimizing sentence length.
• A base corpus of approximately 27K sentences is included under ./data which you reading list will be sourced from.
• A list of diphones in IPA format is also found under ./data.
• example.py contains examples of how to retrieve the corpus and diphones.
• Note: The diphone coverage of list.tsv is more than 1.0. This is because list.tsv contains possibly wrong phonetic predictions and therefore impossible diphones that don't appear in dps.tsv.
• Your reported coverage will be the one returned from tools.coverage(). See example.py for usage.
• The text is in Icelandic and this is the text you will be reading yourself in the upcoming assignments. Some of you are not native Icelandic speakers however. This doesn't mean however that you shouldn't do the assignments in Icelandic and we even think that these projects might provide very interesting results! If you prefer not to complete the assignments in Icelandic, please contact Atli on Teams.