Sawtone v4

Overview

Sawtone v4 preserves word-form evidence across spelling, sound, script, and tokenization. The report positions Sawtone among tokenizer-backed, character-level, and byte-level methods, then evaluates the representation from isolated word geometry through contextual tasks and cross-script transliteration.

Key Findings

• Sawtone keeps spelling, typo, and several cross-script variants close at the whole-word level
• The contextual stack leads DistilUSE on several form-sensitive and cross-script tests, while DistilUSE remains stronger on broad sentence similarity and passage retrieval
• Sawtone selection improves normalized exact match from 40.8% to 54.6% on the Moroccan Arabic evaluation
• Across 3,000 additional predictions, exact match improves from 40.7% to 52.1% and character error falls from 0.259 to 0.205
• Pinyin-to-Hanzi remains outside the reliable form-only regime and requires lexical or semantic disambiguation

Resources

• Full Technical Report (PDF)
• Sawtone model page
• Peer-reviewed Sawtone paper