Tutorial 5: Pitch extraction

The main objective of this tutorial is to perform pitch extraction on the enriched polyglot corpus we used in the previous four tutorials.

As in the other tutorials, import statements and the corpus name (as it is stored in pgdb) must be set for the code in this tutorial to be runnable. The example given below continues to make use of the “tutorial-subset” corpus we have been using in tutorials 1-3.

from polyglotdb import CorpusContext

# corpus_root = './data/LibriSpeech-aligned/'
# corpus_name = 'tutorial'
# export_path = './results/tutorial_4_formants.csv')
corpus_root = './data/LibriSpeech-aligned-subset/'
corpus_name = 'tutorial-subset'
export_path = './results/tutorial_5_pitch.csv')

Vowel phoneme enrichment

In order to analyze pitch, vowel information needs to be encoded in our corpus. See the “Vowel phoneme enrichment” section in Tutorial 4: Vowel formant analysis to encode the required vowel information. If you have already completed tutorial 4, it is not necessary to repeat vowel encoding, and you can move on to the next step.

Pitch Encoding

To extract pitch tracks from the data, we first encode syllable count per word.

with CorpusContext(corpus_name) as c:
c.encode_count('word', 'syllable', 'num_syllables')

Then, the CorpusContext method analyze_pitch is used with a configurable pitch analysis exectuble. Like tutorial 4, in this case, we use praat:

with CorpusContext(corpus_name) as c:
c.config.praat_path = "/usr/bin/praat"
metadata = c.analyze_pitch(algorithm='speaker_adapted', call_back=print)

Pitch is now encoded in all relevant vowels. The next step is to query the data for export.

Exporting a CSV file

We can now query the results using a similar set of commands as in the previous tutorials:

with CorpusContext(corpus_name) as c:
  ## phone comes at beginning of utterance
  q = c.query_graph(c.phone).filter(c.phone.word.begin == c.phone.word.utterance.begin)

  ## restrict just to phone = vowels
  q = q.filter(c.phone.label.in_(vowel_set))

  ## preceding phone is at beginning of the word
  q = q.filter(c.phone.previous.begin == c.phone.word.begin)

  q = q.columns(c.phone.id.column_name('traj_id'),
  results = q.all()

The CSV file generated will then be ready to open in other programs or in R for data analysis. You can see a full version of the script and its expected output when run on the ‘LibriSpeech-subset’ corpora.

Next steps

At this point, the corpus is ready for pitch analysis using R. We have provided an rmd script showcasing a possible approach. We have also provided results for running this script on the “LibriSpeech-aligned” (the full dataset) in a follow-up analysis html. These results contains many more speakers than the subset we have been using in tutorials so far to provide sufficient data for coherent analysis.

See Tutorial 4: Vowel formant analysis for another practical example of interesting linguistic analysis that can be peformed on enriched corpora using python and R. You can also see the `related ISCAN tutorial`_ for R code on visualizing and analyzing the exported results.