Kokoro 82M German (Martin) — CoreML
kikiri-tts/kikiri-german-martin — the German male voice Martin Harbecke, a StyleTTS2 stage-2 fine-tune of Kokoro-82M by kikiri-tts — converted to the same end-to-end CoreML model as aufklarer/Kokoro-82M-CoreML for on-device German TTS on Apple silicon (Neural Engine).
German is not a voice of the multilingual Kokoro bundle: each German voice is a separate weight set trained on German phonemes only, so this bundle carries exactly one voice. The other voice, Victoria, is published as aufklarer/Kokoro-82M-German-Victoria-CoreML.
Model
| Property | Value |
|---|---|
| Parameters | 82M |
| Format | CoreML (.mlmodelc, end-to-end, fixed 128 phonemes / 5 s) |
| Precision | FP16 (Neural Engine) |
| Sample rate | 24 kHz |
| Voice | dm_martin (male) |
| Language | German (de) |
Files
| File | Description |
|---|---|
kokoro_5s.mlmodelc/ |
Pre-compiled end-to-end model (BERT → duration → alignment → prosody → iSTFTNet) |
pipeline_config.json |
Frame bucket / tensor shapes |
vocab_index.json |
The upstream hexgrad IPA → token table |
dict_de.json |
German pronunciation dictionary (49,578 entries) in the voice's phoneme convention |
voices/dm_martin.json |
Voice embedding ({"embedding": [256 floats]}) |
Usage
import KokoroTTS
let tts = try await KokoroTTSModel.fromPretrained(modelId: "aufklarer/Kokoro-82M-German-Martin-CoreML")
let audio = try tts.synthesize(text: "Am 3. Oktober um 18:30 Uhr kostet es 12,80 €.",
voice: "dm_martin", language: "de")
speech kokoro --model aufklarer/Kokoro-82M-German-Martin-CoreML --language de \
"Guten Tag, wie geht es Ihnen?" --output guten-tag.wav
The voice was trained on espeak-ng 1.52 German phonemes as post-processed by
misaki (the kikiri-tts front-end); speech-swift's GermanPhonemizer reproduces
that convention without espeak (number/date/currency normalization,
dict_de.json, inflection and compound fallbacks, rule-based G2P).
Source
Converted from kikiri-tts/kikiri-german-martin
(revision 1e9dcd16ed48, Apache 2.0) with the Soniqo end-to-end Kokoro
CoreML export (FluidInference-aligned SineGen, masked AdaIN) and compiled to
.mlmodelc. Voice embedding: row 128 of the published [510, 1, 256] voicepack.
Links
- speech-swift — Apple SDK
- Docs — install and CLI docs
- soniqo.audio — website
- blog — blog