Phonemizer language coverage

dengjen-tts phonemizes text through three dedicated crates under crates/text/ (dengjen-espeak-phonemizer, dengjen-hebrew-phonemizer, dengjen-pinyin-phonemizer) that Piper, Kokoro, and MeloTTS each call directly rather than through a shared Phonemizer trait (see Architecture for why). Language support is therefore a property of which phonemizer a voice manifest selects, not a single project-wide language list.

What’s actually wired up today

Backend Phonemizer Language coverage

Piper

phoneme_type: "espeak" (default)

Whatever espeak-ng voice the manifest’s espeak.voice field names — see the full list below.

Piper

phoneme_type: "text"

None — raw passthrough, for voices already trained on phonemic/symbol text.

Piper

phoneme_type: "hebrew"

Hebrew only (Nakdimon diacritization + IPA; needs the hebrew Cargo feature).

Piper

phoneme_type: "pinyin"

Mandarin Chinese only.

Kokoro

Derived from the selected voice’s name prefix, not manifest-configurable

en-US, en-GB, Japanese, Mandarin, Spanish, French, Hindi, Italian or Brazilian Portuguese.

MeloTTS

{"type": "espeak", "voice": "…​"}

English, Spanish, French, Japanese, Korean.

MeloTTS

{"type": "pinyin", "model_dir": "…​"}

Mandarin Chinese, with real tone extraction (needs the pinyin Cargo feature).

MeloTTS’s pinyin path and Piper’s hebrew/pinyin paths are the only ones with dedicated, tone- or diacritic-aware handling; espeak-backed paths get IPA phonemes only.

espeak-ng’s full language list

Piper’s espeak phoneme type and MeloTTS’s espeak phonemizer both accept any language/locale code espeak-ng itself ships. The list below is extracted directly from dengjen-espeak-rs-sys’s bundled `espeak-ng-data/lang/ (the actual fork this project vendors), not from upstream espeak-ng documentation, so it reflects what’s really available at runtime:

ab, af, am, an, ar, as, az, ba, be, bg, bn, bpy, bs, ca, ca-ba, ca-nw, ca-va, chr, cmn,
cmn-Latn-pinyin, crh, cs, cv, cy, da, de, el, en, en-029, en-GB-scotland, en-GB-x-gbclan,
en-GB-x-gbcwmd, en-GB-x-rp, en-Shaw, en-US, en-US-nyc, eo, es, es-419, et, eu, fa, fa-Latn,
fi, fo, fr, fr-BE, fr-CH, ga, gd, gn, grc, gu, hak, haw, he, hi, hr, ht, hu, hy, hyw, ia, id,
io, is, it, ja, jbo, ka, kaa, kk, kl, kn, ko, kok, ku, ky, la, lb, lfn, lt, ltg, lv, mi, mk,
ml, mn, mr, ms, mt, mto, my, nb, nci, ne, nl, nog, om, or, pa, pap, piqd, pl, ps, pt, pt-BR,
py, qdb, qu, quc, qya, ro, ru, ru-cl, ru-LV, rup, sd, shn, si, sjn, sk, sl, smj, sq, sr, sv,
sw, ta, te, th, ti, tk, tn, tr, tt, ug, uk, ur, uz, vi, vi-VN-x-central, vi-VN-x-south, xex,
yue, yue-Latn-jyutping

This is espeak-ng’s own capability, not a dengjen-tts guarantee. dengjen-tts’s own test suite only exercises English voices end-to-end; any other language on this list will phonemize (espeak-ng supports it), but whether a given Piper or MeloTTS ONNX voice — trained on some specific language’s phoneme set — actually produces intelligible speech from that phonemization is untested by this project and depends entirely on the voice model itself.

Japanese phonemizer for Kokoro (optional)

Kokoro was trained on misaki phonemes, and eSpeak’s Japanese output approximates them poorly. The japanese Cargo feature on dengjen-tts-kokoro (forwarded as --features japanese by the CLI, gRPC, C API and Python frontends) phonemizes j* presets with jpreprocess, a Rust rewrite of OpenJTalk, and misaki’s katakana-to-phoneme table. Without the feature, j* presets use eSpeak.

The feature needs a jpreprocess-format NAIST-jdic dictionary, which the binary does not bundle. Download naist-jdic-jpreprocess.tar.gz from the jpreprocess v0.15.0 release (the dictionary version must match the crate version), extract it, and set DENGJEN_JA_DICT_DIR to the extracted directory. With the variable unset, a Kokoro config that lists j* voices still loads, and synthesizing with a j* preset fails with an error naming the variable; other presets are unaffected. A variable pointing at a directory that is not a valid, matching dictionary fails the load with an error naming the variable.

Symbols that Kokoro’s vocabulary lacks are remapped (g to ɡ, palatalized consonants to consonant + ʲ, labialized kw/gw to kᵝ/ɡᵝ), because the tokenizer drops unknown symbols silently. Pitch accent is not encoded in the phonemes, matching misaki.