cryptomilk/kokoro-german-kerstin

Kokoro German — Kerstin (df_kerstin)

A German female voice for Kokoro-82M. Trained with kikiri-tts on the Kerstin dataset.

Use with Crane

Use voices/df_kerstin.bin. Copy it into Crane's voice directory of Kokoro. Crane picks it up automatically.

Use with Python

Use voices/df_kerstin.pt.

uv run kokoro --text "Hallo, ich bin Kerstin." -o output.wav -l d --voice df_kerstin

Or:

from kokoro import KModel, KPipeline

model = KModel()
pipeline = KPipeline(lang_code="d", model=model)
for result in pipeline("Hallo, ich bin Kerstin.", voice="voices/df_kerstin.pt"):
    print(result.audio)

Developer files

This repo also includes the raw training checkpoints:

• data/kikiri_german_kerstin_stage1_ep10.pth — Stage 1 checkpoint. Decoder, alignment, and vocoder. Final epoch.
• data/kikiri_german_kerstin_stage2_ep10.pth — Stage 2 checkpoint. Prosody predictor, trained with a WavLM adversarial loss. Final epoch. The voices/ files were extracted from this checkpoint.

Most users don't need these. They're here for continuing training, extracting a different voicepack variant, or inspecting the model.

License

Apache 2.0. Matches Kokoro-82M and kikiri-tts. The Kerstin training dataset is CC0-1.0.