phonk 0.0.3

Some helpful phonetics and phonology functions

phonk: some helpful phonetics and phonology functions

Pronounced [fʌŋk], [fɑŋk], [fɔŋk], or even [foʊŋk].

Installation

Run the following:

pip intall phonk

Usage

RhymeComp

So far, the only module is rhymescore, which contains the RhymeComp class. This is a tool for finding and investigating (poetic) rhymes. Import the RhymeComp class directly.

from phonk.rhymescore import RhymeComp

RhymeComp('lives','read').get_score()

The get_score() method returns a non-negative integer. 0 indicates the rhymes are identical. Any positive number is the number of featural changes between the two rhymes. (See below for exact calculation method.)

To see verbose output (that explains the exact featural differences), use the print_details() method. This method takes an integer argument (default is verbose=1); the higher the number, the more detail shown.

RhymeComp('lives','read').print_details()

The above will output the following:

'lives' vs. 'read'       SCORE: 10
Segment Comparisons
---
IH1     EH1     1
V       None    6
Z       D       3

Feature Comparisons
---
       IH   EH
high  1.0  0.0
                   Z    D
continuant       1.0  0.0
delayed release  1.0  0.0
strident         1.0  0.0

Possible Pronunciation Comparison:
---
('IH1', 'V', 'Z')       ('EH1', None, 'D')      10
('IH1', 'V', 'Z')       ('IY1', None, 'D')      10
('AY1', 'V', 'Z')       ('EH1', None, 'D')      12
('AY1', 'V', 'Z')       ('IY1', None, 'D')      14

The function has pulled each word from the CMU pronouncing dictionary. Because each word has multiple listed pronunciations, and the rhymes are of different length, the function compares all possible alignments of all possible pronunciations. The best alignment of each pronunciation is printed at the end. The Segment Comparison shows that the lower score occurs when Z is compared with D. The V of lives is compared with an empty position. The default score for such a comparison is 6; this can be changed by setting a value to e_comp when creating the object.

The details of each segment-to-segment comparison are shown in the Feature Comparison tables. The feature set is based on Hayes 2012. This is a relatively standard set of features; definitions for which can be found in chapter 4 of the aforementioned textbook. To access these features directly, call the differing_features() method, which returns them in a list.

print(RhymeComp('lives','read').differing_features())

>>> ['high', 'continuant', 'delayed release', 'strident']

Under the hood

Getting pronunciations

The function will load and search the CMU pronouncing dictionary via cmudict for pronunciations of the given words. As the CMU might not be accurate, the function can also accept an ARPAbet list representation directly.

If there are multiple pronunciations found, the lowest-scoring pronunciation is used for the base calculation. Alternate pronunciations can be viewed via the print_details(verbose=3) method, or the possible_prons attribute directly.

Defining 'rhyme'

By default, the rhyme comparison looks at the last VC*# sequence of the two words (i.e. the linguistic rhyme of the final syllable). If you want to compare multiple syllables, set syllables=n upon object creation. If a given word is shorter than n syllables, the comparison stops at the length of the shortest word. Currently, stress information is ignored for rhyme calculation.

Aligning rhymes

If two rhymes are of different length, they must be aligned. The shorter of the two rhymes is padded with an empty object (None) at every possible position, and each of these sequences is compared with the longer rhyme. The lowest-scoring comparison is used for the eventual result. Any comparison of a segment to an empty position is given an arbitrary score of 6. This can be changed by setting a different value to e_comp upon object creation.

Scoring rhymes

The score of the rhyme comparison is the sum of all featural differences between the two segments, in addition to the e_comp value for every empty comparison done. In essence, a fairly standard faithfulness calculation is done between the two strings. In the lives vs read example above, the score breakdown is as follows:

Segment Comparisons
---
IH1     EH1     1
V       None    6
Z       D       3

Since the rhyme for read was shorter, it was padded with None. The placement of this padding is such that the score is minimized; comparing Z with D and V with None results in a lower score than if e.g. V is compared with D and Z with None.

Once there are two given sequences to compare, score lookup is relatively fast. The diffs.csv table includes all ARBAbet symbols as rows and columns with the number of featural differences between them as the values. This table is loaded as a pandas dataframe and called for every segmental comparison.

Only the first instance call of RhymeComp loads the relevant tables into memory, making subsequent calculations relatively fast. (h/t Anupam Basu)

Acknowledgements

TK