A native C# port of misaki — Kokoro TTS's official G2P engine — with zero Python and zero espeak in the loop.
Every module is verified at mass scale against the reference implementation, on held-out corpora it was never tuned on:
| Module | Verified against | Parity |
|---|---|---|
ChineseG2P (jieba posseg + tone sandhi + pinyin→zhuyin + cn2an) |
misaki 0.9.4 ZHG2P(version='1.1') |
100.00% (1936/1936 sentences) |
EnglishG2P (lexicon, stress, numbers, currency, homographs) |
misaki 0.9.4 en.G2P |
51,623 fixtures at 100% |
EnglishTokenizer |
spaCy en_core_web_sm tokenization |
100% (56,216 lines) |
EnglishTagger (2.8MB perceptron distilled from spaCy) |
spaCy Penn Treebank tags | 97.5% tags → 99.20% end-to-end phoneme parity |
var english = new EnglishG2P(EnglishG2P.DefaultTagger); // American English (british: true for en-GB)
var (phonemes, tokens) = english.Phonemize("On March 3rd, I read that the record was $1,234.56!");
var zhuyin = ChineseG2P.Phonemize("你好，世界！欢迎使用。"); // Kokoro v1.1 format, tone sandhi included
var arrows = ChineseG2P.PhonemizeLegacy("你好，世界！"); // Kokoro v1.0 format, for the older zh voicesPhoneme output uses Kokoro's phoneme alphabet. Rare out-of-lexicon English words return unk (❓) unless you plug an espeakFallback; latin spans inside Chinese text route through ChineseG2P.EnglishPhonemizer when set.
The building blocks are public and useful on their own: EnglishTokenizer (spaCy-equivalent tokenization), EnglishTagger (fast PTB tagging), PossegTagger (bit-exact jieba posseg / cut-for-search), ToneSandhi, and ZhFrontend.
The dictionaries (~22MB) embed into the assembly — pinyin/jieba tables, misaki's English lexicons, spaCy-derived tokenization rules, and the distilled tagger. Everything regenerates from generators/*.py against the upstream sources.
- Apache-2.0, like misaki itself. See NOTICE for the full lineage (PaddleSpeech, jieba, pypinyin, spaCy, num2words).
- Versioning tracks the upstream misaki release the port is verified against (currently 0.9.4).