Phonetic Decomposition¶

This section demonstrates how to break graphemes down to their phonetic building blocks and rebuild them.

What is Decomposition?¶

Every Indic grapheme cluster is built from a consonant (the base shape) and a vowel (which determines how the consonant sounds). For example:

கா = consonant க + vowel ஆ
கி = consonant க + vowel இ

Decomposition splits a grapheme into these two parts, represented as: - The consonant with a hal/virama marker (the silent consonant form) - The base independent vowel

For example: கா decomposes into க் + ஆ.

Decomposing Tamil Text¶

from grapheme_kit import decompose

# "ka" with the "aa" vowel
print(decompose("கா"))
# Output: க்ஆ

# A full word — vowels are separated from their consonants
print(decompose("வணக்கம்"))
# Output: வ்அண்அக்க்அம்

Decomposing Sinhala Text¶

from grapheme_kit import decompose

print(decompose("සිංහල"))
# Output: ස්ඉංහ්අල්අ

Composing Back¶

Applying compose() to a decomposed string will reconstruct the original grapheme clusters.

from grapheme_kit import decompose, compose

original = "வணக்கம்"
decomposed = decompose(original)
restored = compose(decomposed)

print(restored)           # Output: வணக்கம்
assert original == restored  # Always true

Practical Use: Changing a Vowel¶

Decomposition is particularly useful when you want to programmatically change the vowel of a consonant.

from grapheme_kit import decompose, compose

# Original: "ka" (க + ஆ)
# Goal: change the vowel to "இ" to get "ki" (கி)

decomposed = decompose("கா")
# decomposed is now "க்ஆ"

# Replace the vowel part manually
modified = "க்இ"

result = compose(modified)
print(result)
# Output: கி

Summary¶

You have completed the tutorial series. Here is what you covered:

The difference between code points and graphemes, and why it matters for evaluation.
Why grapheme-aware metrics complement -- rather than replace -- chrF, chrF++, CER, and CharBLEU, and produce identical results for scripts where code points and graphemes already coincide.
How the Graphemizer pipeline works internally.
How to segment Tamil conjuncts and Sinhala ZWJ sequences correctly.
How to decompose and compose graphemes for phonetic manipulation.

Continue to: NLP Evaluation for practical task-specific recipes.