Skip to content

g2p

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

g2p

README.md

piper-plus-g2p

Multilingual G2P (Grapheme-to-Phoneme) for TTS. eSpeak-ng free. MIT licensed. 8 languages.

Standalone package — piper-plus TTS エンジンなしで G2P 単体として利用可能です。任意の TTS エンジンと組み合わせて使えます。

Why piper-plus-g2p?

  • MIT licensed -- no eSpeak-ng (GPL) dependency in your TTS pipeline
  • 8 languages -- JA, EN, ZH, KO, ES, FR, PT, SV with consistent IPA output
  • IPA-first design -- returns pure IPA token sequences, ready for any TTS model

Comparison

piper-plus-g2p phonemizer gruut Misaki
License MIT GPL (eSpeak-ng) MIT Apache-2.0
Languages 8 100+ 20+ EN only
eSpeak-ng required No Yes No No
IPA output Yes Yes Yes Yes

Installation

pip install "piper-plus-g2p==0.3.1"       # Rule-based languages (ES, FR, PT, SV)
pip install "piper-plus-g2p[ja,en]==0.3.1" # Japanese + English
pip install "piper-plus-g2p[all]==0.3.1"   # All 8 languages

Note: The ja extra requires pyopenjtalk-plus, which provides pre-built wheels for Linux, macOS, and Windows. See pyopenjtalk-plus for platform details.

Quick Start

from piper_plus_g2p import get_phonemizer

ja = get_phonemizer("ja")
ja.phonemize("こんにちは")
# -> ["k", "o", "[", "N_n", "n", "i", "ch", "i", "w", "a"]

en = get_phonemizer("en")
en.phonemize("Hello world")
# -> ["h", "ə", "l", "ˈ", "o", "ʊ", " ", "w", "ˈ", "ɜ", "ː", "l", "d"]

Supported Languages

Language Code Extra Backend Notes
Japanese ja piper-plus-g2p[ja] pyopenjtalk-plus Context-dependent N variants, prosody info
English en piper-plus-g2p[en] Bundled g2p-en 2.1.0 source CMU-dict + neural fallback; Apache-2.0 / MIT notices included
Chinese zh piper-plus-g2p[zh] pypinyin Pinyin-to-IPA, ZH-EN code-switching (loanword pinyinisation)
Korean ko piper-plus-g2p[ko] g2pk2 Optional dependency
Spanish es -- Rule-based No external dependency
French fr -- Rule-based No external dependency
Portuguese pt -- Rule-based No external dependency
Swedish sv -- Rule-based No external dependency

Advanced Usage

Multilingual (Composite Language Code)

Pass a hyphen-joined code like "ja-en-zh" to get_phonemizer to automatically create a MultilingualPhonemizer. Language detection is Unicode-based, so mixed-script text is handled without explicit tagging.

from piper_plus_g2p import get_phonemizer

multi = get_phonemizer("ja-en-zh")
tokens = multi.phonemize("こんにちは Hello 你好")

Version 0.3.1 bundles the reviewed English backend under a private namespace; the en extra installs NumPy, NLTK, joblib and inflect, without installing the external g2p-en distribution or its unused GPL Distance dependency. The vendored namespace is an implementation detail, not a public import path. The already published piper-plus-g2p==0.3.0 wheel still has the old dependency metadata. See the migration guide and third-party notices.

ZH-EN Code-Switching (English Inside Chinese Context)

When an English token (acronym, brand, technical term) appears next to a Chinese segment, MultilingualPhonemizer routes it through ChinesePhonemizer.phonemize_embedded_english so it is rendered as Mandarin pinyin instead of US English -- matching how Chinese speakers naturally pronounce inserted English words.

from piper_plus_g2p import get_phonemizer

multi = get_phonemizer("zh-en")
multi.phonemize("请打开 GPS")          # GPS -> ji4-pi4-ai1-si4 (pinyin)
multi.phonemize("我喜欢用 Python 写代码")  # Python -> pai4-sen1
multi.phonemize("Hello world")        # No zh context -> bundled English backend

Custom overrides (acronym/loanword/letter_fallback) are loaded from a JSON file shaped like the bundled data/zh_en_loanword.json:

from piper_plus_g2p.chinese import ChinesePhonemizer

zh = ChinesePhonemizer(zh_en_loanword_dict_paths="my_overrides.json")
zh.phonemize_embedded_english("MYORG")

PiperEncoder

Converts IPA token lists into integer phoneme_ids for Piper ONNX models.

from piper_plus_g2p.encode import PiperEncoder, get_phoneme_id_map

id_map = get_phoneme_id_map("ja")
encoder = PiperEncoder(id_map)
phoneme_ids = encoder.encode(["k", "o", "[", "N_n", "n", "i", "ch", "i", "w", "a"])

Piper Model Compatibility

piper-plus-g2p produces phoneme tokens directly compatible with Piper TTS ONNX models. Use PiperEncoder with the model's phoneme_id_map from config.json:

import json
from piper_plus_g2p import get_phonemizer
from piper_plus_g2p.encode import PiperEncoder

with open("model/config.json") as f:
    config = json.load(f)

encoder = PiperEncoder(config["phoneme_id_map"])
phonemizer = get_phonemizer("ja")
tokens = phonemizer.phonemize("こんにちは")
phoneme_ids = encoder.encode(tokens)
# -> ready for ONNX inference

Use strict=True to raise errors on unknown tokens instead of silently skipping them:

encoder = PiperEncoder(config["phoneme_id_map"], strict=True)

Cross-Platform Consistency

piper-plus-g2p is also available as:

  • Rust crate: piper-plus-g2p on crates.io
  • Go module: github.com/ayutaz/piper-plus/src/go/phonemize
  • npm package: @piper-plus/g2p for browser/WASM

All three implementations share the same PUA mapping table and are validated against a common test fixture.

Requirements

License

MIT