Kokoro 82M — German "Martin"
On-device German text-to-speech, packaged for iOS. The same voice is provided in two runtimes: ONNX for CPU inference and Core ML for the Apple Neural Engine.
Contents
| File | Runtime | Size | Notes |
|---|---|---|---|
kokoro-martin.onnx |
ONNX | 310 MB | Full-precision model for the CPU path |
voices-martin.npz |
ONNX | 0.5 MB | Style tensor for the "Martin" voice |
kokoro-german-martin-ios26-100k-lexicon.tar.gz |
Core ML | 71 MB | Current package. Profile ios18-static-bucket, no palettization, iOS 18+ |
kokoro-german-martin-ios26-release-parity.tar.gz |
Core ML | 78 MB | Previous package, kept as a fallback. Profile ios26-release-parity, palettized, iOS 26+ |
vocab_index.json |
both | — | Token index (114 code points) |
phonemizer/de_lexicon.json |
both | 36 MB | German pronunciation lexicon |
phonemizer/LICENSES.md |
— | — | Licences of the lexicon sources |
The current Core ML archive is self-contained: it bundles vocab_index.json, the converted voice
(voices/martin.bin) and the phonemizer alongside the two compiled models.
Why the current package requires less, not more
The ios26-release-parity build uses palettized weights (196 constexprLutToDense operations),
which are only available from the iOS 26 runtime. The ios18-static-bucket build has none, runs
from iOS 18, and is the one verified on device — including an iPhone 12 Pro (A14), where the
palettized build's runtime floor would have excluded it.
Core ML package layout
artifact_manifest.json profile, checksums, toolchain versions
kokoro_frontend.mlmodelc/ predictor (CPU)
kokoro_backend.mlmodelc/ decoder (ANE-eligible)
vocab_index.json
voices/martin.bin
phonemizer/
artifact_manifest.json is the authority on what a given archive actually is — filenames are
kept stable for client compatibility and do not always describe their contents.
Provenance
- Voice weights:
kikiri-tts/kikiri-german-martin— a StyleTTS2 stage-2 fine-tune (speaker: Martin Harbecke) - German base:
dida-80b/kokoro-german-hui-multispeaker-base - Architecture:
hexgrad/Kokoro-82M - Core ML export recipe:
Jud/kokoro-coreml@ 0.11.0
This repository repackages those weights for on-device use. It contains no new training.
Licence
Apache-2.0, inherited from the upstream weights and the export tooling. The pronunciation
lexicon carries its own attributions in phonemizer/LICENSES.md.
Used by
Enclavi — a private, on-device notes and meeting assistant for iOS, which downloads these packages at runtime.