Phonetic Similarity Analysis of Indonesian Text: Comparing Custom Soundex, Jaro-Winkler, Cosine Similarity, and Euclidean Distance

Indonesian text exhibits rich phonetic variation: regional spelling variants, old-to-new orthography changes such as tj→c and oe→u, and digraphs such as ng, ny, kh, and sy that standard phonetic algorithms designed for English fail to capture. This paper presents a comparative analysis of four phonetic similarity algorithms on Indonesian word pairs: (1) a Custom Soundex adapted to Indonesian phonology with 5-digit codes, digraph-aware encoding, old-spelling normalization, and silent-h stripping; (2) Jaro-Winkler similarity computed on grapheme-to-phoneme (G2P) representations, implemented as phoneme similarity via soft Levenshtein distance with phoneme-class substitution costs; (3) Cosine Similarity over positional phoneme vectors with a length-ratio penalty; and (4) Euclidean Distance over the same positional vectors. The evaluation uses a dataset of 10,000 Indonesian word pairs labeled Mirip (similar) or Beda (different), balanced 5,000/5,000 across phonetically similar and control pairs. Decision thresholds were tuned per method: identical codes for Soundex, 0.80 for Jaro-Winkler+G2P, 0.65 for Cosine Similarity, and a distance of 1.5 for Euclidean. Each algorithm classifies pairs with its tuned decision threshold. Custom Soundex ID achieved the best performance with 99.2% accuracy and an F1-score of 0.992, followed by Cosine Similarity (94.9%, F1 0.948), Jaro-Winkler+G2P (91.9%, F1 0.912), and Euclidean Distance (91.7%, F1 0.910). Confusion matrix analysis shows that phoneme-aware encoding resolves most false positives caused by coincidental prefix overlap and subword absorption. The main contribution is empirical evidence that a lightweight, linguistically informed phonetic code outperforms generic string similarity measures for Indonesian, together with a documented preprocessing pipeline of orthography normalization, G2P conversion, and silent-h stripping that is reusable for Indonesian fuzzy text matching.