anchpop/lexide-parsley
Updated
Multilingual NLP + speech models: tokenization, lemmatization, POS, dependency parsing, and phonetic transcription.
Note Production tagger — XLM-RoBERTa encoder + byte-level minGRU char tokenizer. Runs tokenization, POS, lemmatization, and dependency tagging on CPU in real time.
Note The teacher — a 31B Gemma-4 LoRA fine-tune that generates the silver training data and gold labels parsley learns from (and serves as a fallback for harder languages).
Note Multilingual phonetic transcription: speech → narrow IPA of what was actually pronounced. Powers pronunciation feedback for language learners.