LipSpell
A language-independent phonetic spelling correction framework that builds a phoneme distance matrix from IPA distinctive features, combines keyboard geometry and corpus frequency for weighted scoring, and accelerates candidate retrieval through two-stage indexing, supporting cross-lingual transfer.
Timeline (last 90 days)
LipSpell uses a language-agnostic phoneme scoring engine paired with a replaceable language front-end, pulling spelling correction from the orthographic level to the phonological level.
LipSpell's phoneme distance matrix is precomputed based on IPA distinctive features, and the closer two phonemes are in feature space, the lower the substitution cost.
LipSpell's unified scoring function combines three components: phoneme similarity, keyboard geometric distance, and a corpus unigram frequency prior.
LipSpell uses two-stage symmetric deletion plus a seven-class phoneme-encoding index to accelerate candidate recall, doing a coarse filter first and then fine ranking.
On six English benchmarks, LipSpell comprehensively outperforms Norvig and Hunspell.
On the Norvig1 benchmark, LipSpell achieves 78.52% accuracy, while Aspell achieves 75.56%.
On the Holbrook benchmark, LipSpell achieves 32.00% accuracy, while Aspell achieves 26.67%.
When the fixed weights are transferred directly to Brazilian Portuguese, LipSpell achieves 73.00% Top-1 on 500 Wikipedia samples (without NFD tolerance), and 75.80% after tolerating NFD.
Under tuned conditions, LipSpell's gains over Norvig are relatively pronounced on Portuguese (86.97% vs 76.75%), Turkish (81.67% vs 79.05%), Hebrew consonantal script (74.59% vs 62.30%), and French (74.53% vs 67.92%); English (80.37%) and Russian (85.67%) remain optimal with fixed weights.
LipSpell was evaluated on six languages, covering five branches of the Germanic, Romance, Slavic, Semitic, and Turkic families, spanning the Indo-European, Afro-Asiatic, and Turkic language families and three writing systems.
1 more timeline events
All Facts (11)
LipSpell uses a language-agnostic phoneme scoring engine paired with a replaceable language front-end, pulling spelling correction from the orthographic level to the phonological level.
50%UnverifiedLipSpell's phoneme distance matrix is precomputed based on IPA distinctive features, and the closer two phonemes are in feature space, the lower the substitution cost.
50%UnverifiedLipSpell's unified scoring function combines three components: phoneme similarity, keyboard geometric distance, and a corpus unigram frequency prior.
50%UnverifiedLipSpell uses two-stage symmetric deletion plus a seven-class phoneme-encoding index to accelerate candidate recall, doing a coarse filter first and then fine ranking.
50%UnverifiedOn six English benchmarks, LipSpell comprehensively outperforms Norvig and Hunspell.
50%UnverifiedOn the Norvig1 benchmark, LipSpell achieves 78.52% accuracy, while Aspell achieves 75.56%.
50%UnverifiedOn the Holbrook benchmark, LipSpell achieves 32.00% accuracy, while Aspell achieves 26.67%.
50%UnverifiedWhen the fixed weights are transferred directly to Brazilian Portuguese, LipSpell achieves 73.00% Top-1 on 500 Wikipedia samples (without NFD tolerance), and 75.80% after tolerating NFD.
50%UnverifiedUnder tuned conditions, LipSpell's gains over Norvig are relatively pronounced on Portuguese (86.97% vs 76.75%), Turkish (81.67% vs 79.05%), Hebrew consonantal script (74.59% vs 62.30%), and French (74.53% vs 67.92%); English (80.37%) and Russian (85.67%) remain optimal with fixed weights.
50%UnverifiedLipSpell was evaluated on six languages, covering five branches of the Germanic, Romance, Slavic, Semitic, and Turkic families, spanning the Indo-European, Afro-Asiatic, and Turkic language families and three writing systems.
50%UnverifiedThe phoneme front-end's dependence on G2P quality for low-resource languages is an unavoidable constraint when extending LipSpell to more languages.
50%