feat: add German (de) language support

[semidark]

Summary

Register German as a supported language in Kokoro's pipeline, CLI, and kokoro.js voice list.

Changes

• Add 'de': 'd' to ALIASES and d='de' to LANG_CODES in pipeline.py
• Add a dedicated elif lang_code == 'd': branch that uses misaki.de.DEG2P() instead of the generic espeak fallback
• Add "d" (German) to the CLI language choices in __main__.py
• Add commented-out German voice stubs (df_anna, dm_daniel) to kokoro.js/src/voices.js

German follows the same dedicated-G2P pattern as Japanese (JAG2P) and Chinese (ZHG2P).

See also:

• misaki PR for German G2P: feat: add German G2P with text normalization misaki#97
• Related bugfix PR: fix: remove stale chunking warning for espeak languages #318

Usage

This PR enables lang_code='d' / lang_code='de' in upstream Kokoro's pipeline and CLI.

pipeline = KPipeline(lang_code='d', repo_id='...')
# or equivalently:
pipeline = KPipeline(lang_code='de', repo_id='...')

German synthesis depends on a compatible German-trained model family and matching voice assets.

Context: Training Recipe and German Model Work

This PR comes out of kokoro-deutsch, a community project that reverse-engineered a working Kokoro fine-tuning pipeline using a patched StyleTTS2. The repo contains:

• End-to-end training recipe (dataset prep -> Stage 1 -> Stage 2 -> voicepack extraction)
• Patched StyleTTS2 submodule with critical fixes for stable Stage 2 training
• A forked misaki with German text normalization (misaki.de.DEG2P)
• Full documentation (training guide, troubleshooting, architecture notes)

A German multi-speaker base model has been published by the community:

• Model: dida-80b/kokoro-deutsch-hui-base
• Dataset: dida-80b/hui-german-51speakers
• Intermediate inference sample

Important technical note: German voicepacks extracted from this work are tied to the German-trained model family, not to upstream hexgrad/Kokoro-82M. Because of that, this PR is intentionally limited to upstream language enablement, while the trained German model assets currently live in the external community project.

Timeline

• April 10: Published kokoro-deutsch
• April 10-14: Fixed critical Stage 2 bugs
• April 11: Community contributors began training runs
• April 14: Published the initial fork commit adding German language code support
• April 17: Added dedicated German text normalization via misaki.de.DEG2P and opened upstream misaki PR Docker compose error #97

Relates to #290

[semidark]

/claim #290

[semidark]

Status update: Voice .pt files are in progress. We have a trained multi-speaker base model and are currently extracting and validating individual voicepacks. I'll update this PR with at least one German voice file (following the d{f,m}_name naming convention) once quality validation is complete.

Happy to adjust the implementation based on any feedback in the meantime.

[semidark]

Small technical clarification after reviewing model compatibility more carefully:

the German-trained assets coming out of kokoro-deutsch are tied to the German-trained checkpoint family, not automatically to upstream hexgrad/Kokoro-82M. So standalone German voice .pt files are not enough by themselves unless paired with a compatible German-trained model.

Because of that, this PR is intentionally scoped to upstream language enablement and related plumbing:

• register d / de
• remove the stale espeak chunking warning
• add German voice stubs on the JS side

The trained German model work continues in the external community project:
https://github.com/semidark/kokoro-deutsch

If upstream wants German support wired in at the code level first, this PR keeps that change minimal and technically accurate.

[semidark]

Split out the stale chunking warning fix into its own PR (#318) since it's an independent bugfix that benefits all espeak-ng languages. This PR now contains only the German language registration and kokoro.js voice stubs.

[semidark]

Update: German is now wired through a dedicated misaki.de.DEG2P() path instead of the generic EspeakG2P(language='de') fallback.

This matches the architecture used for Japanese (JAG2P) and Chinese (ZHG2P):

• dedicated G2P module in misaki
• language branch in KPipeline
• normalization before phonemization

Upstream misaki PR for the German module is here:
hexgrad/misaki#97

That PR adds German text normalization for numbers, dates, times, currency, and abbreviations before espeak-ng phonemization.

[TheMasterOfDisasters]

@semidark
Hi, I started wiring experimental German support into our KokoroTTS Docker/UI/API fork and wanted to test the Eva-K assets mentioned here.

I was able to find references to dida-80b/kokoro-deutsch-eva-k, and at one point the model page appeared visible, but programmatic access now seems unavailable:

• https://huggingface.co/api/models/dida-80b/kokoro-deutsch-eva-k returns {"error":"Invalid username or password."}
• direct file access to eva_k.pt returns 401 Unauthorized
• searching models for dida-80b + kokoro returns no model results
• I only see the public dataset dida-80b/hui-german-51speakers-synthetic

I would like to test and possibly integrate the German voice path, but without a publicly downloadable Kokoro-compatible model/voicepack I cannot really validate it.

Is dida-80b/kokoro-deutsch-eva-k intentionally private now, renamed, or temporarily unavailable? If there is a new public repo or asset location for kokoro_german_converted.pth and eva_k.pt, I would be happy to try it.