feat: add German G2P with text normalization

[semidark]

Summary

Add German G2P module (misaki/de.py) with text normalization and espeak-ng phonemization.

Changes

• misaki/de.py: DEG2P class + normalize_text_de() function
• tests/test_de.py: 61 unit tests (CI-safe, no espeak needed) + 4 integration tests
• pyproject.toml: add de optional dependency

What normalize_text_de() handles

• Cardinal numbers (42 -> zweiundvierzig)
• Ordinals (3. -> dritte)
• Years (1985 -> neunzehnhundertfünfundachtzig)
• Dates (24.12.2024 -> vierundzwanzigste Dezember zweitausendvierundzwanzig)
• Times (14:30 -> vierzehn Uhr dreißig)
• Currency (€29,99 -> neunundzwanzig Euro und neunundneunzig Cent)
• 30+ German abbreviations (Dr., GmbH, z.B., usw., months, etc.)
• German-format numbers (1.234,56 with thousand dots + decimal comma)
• Quote normalization, whitespace cleanup

Architecture

DEG2P follows the same pattern as KOG2P (Korean) and HEG2P (Hebrew):

class DEG2P:
    def __init__(self):
        from .espeak import EspeakG2P
        self.espeak = EspeakG2P(language='de')

    def __call__(self, text) -> Tuple[str, None]:
        text = normalize_text_de(text)
        return self.espeak(text)

The deferred import allows normalize_text_de() to be tested without phonemizer installed.

Usage

from misaki.de import DEG2P

g2p = DEG2P()
phonemes, _ = g2p('Dr. Müller kaufte am 3. Mai um 14:30 Uhr 3 Pakete für €29,99.')

Context

This module was developed as part of kokoro-deutsch, a community project for fine-tuning Kokoro for German. A trained German multi-speaker base model is published at dida-80b/kokoro-deutsch-hui-base.

No new heavy dependencies are introduced. The de extra only adds phonemizer-fork and espeakng-loader, which are already used by misaki[en].

[apples-kksk]

I found a few small German-normalizer edge cases while testing this PR and opened a follow-up against your branch here: semidark#1

It covers:

• not consuming Uhr inside longer words like Uhrzeit
• leaving invalid times such as 25:00 Uhr / 23:99 Uhr unchanged instead of expanding around the colon
• using Decimal for currency rounding so €9,999 becomes zehn Euro rather than neun Euro und einhundert Cent

Verification:

• local default env: python -m pytest tests/test_de.py -q -> 65 passed, 4 skipped
• isolated Python 3.11 with .[de]: full tests/test_de.py -> 69 passed

I also noticed pyproject.toml adds the de extra, but uv.lock has not been regenerated for it yet. I left that untouched because uv lock currently fails for the existing he extra resolution on Python 3.12/3.13 (mishkal-hebrew>=0.3.2).