Tutorial 5: Pitch extraction

The objective of this tutorial is to perform pitch (f0) extraction on the enriched PolyglotDB corpus we used in Tutorials 1-3.

The complete Python script for Tutorial 5 is available here: tutorial scripts.

Workflow

This tutorial assumes you have already run Tutorials 1 and 2, which import and enrich the corpus with syllable, speaker, and utterance information.

Tutorial 5 can be followed in two ways, like other tutorials (see here):

• Pasting in commands one by one into the Python interpreter

• Running the entire script at once (python tutorial_5.py)

Running the whole script is the usual workflow for PolyglotDB, but the commands are shown one by one here to make it easier to follow along.

As in previous tutorials, import statements and the location of the corpus (corpus_root) must be set for the code in this tutorial to be runnable. (You also need to make sure the directory where you will save the CSV file, here results/, exists.)

import os
import re
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:

Note

When performing analysis with Praat, you might encounter an EOFError due to the use of multiprocessing by polyglotdb. To avoid this, include the statement: if __name__ == '__main__': at the beginning of your program.

with CorpusContext(corpus_name) as c:
c.reset_acoustic_measure('pitch')
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.subset == 'vowel')

  # 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'),
                c.phone.label.column_name('vowel'),
                c.phone.previous.label.column_name('consonant'),
                c.phone.following.label.column_name('following_phone'),
                c.phone.word.label.column_name('word'),
                c.phone.word.duration.column_name('word_duration'),
                c.phone.word.transcription.column_name('word_transcription'),
                c.phone.word.following.transcription.column_name('following_word_transcription'),
                c.phone.begin.column_name('begin'),
                c.phone.end.column_name('end'),
                c.phone.discourse.name.column_name('discourse'),
                c.phone.speaker.name.column_name('speaker'),
                c.phone.speaker.sex.column_name('sex'),
                c.phone.pitch.track.column_name('f0'))

  # Optional: Use order_by to enforce ordering on the output for easier comparison with the sample output.
  q = q.order_by(c.phone.label)
  results = q.all()
  q.to_csv(export_path)

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, compiled here (source) These results were found using the full LibriSpeech-aligned dataset, which contains many more speakers than the subset we have been using in tutorials so far.