cstr/moonshine-tiny-de-dattazigzag-GGUF

Moonshine Tiny (German, dattazigzag) -- GGUF

GGUF conversions and quantisations of dattazigzag/moonshine-tiny-de for use with CrispStrobe/CrispASR.

Available variants

File	Quant	Size	Notes
moonshine-tiny-de-dattazigzag-f16.gguf	F16	52 MB	Half precision
moonshine-tiny-de-dattazigzag-q4_k.gguf	Q4_K	17 MB	Best size/quality tradeoff

Model details

• Architecture: Conv1d stem + 6L transformer encoder + 6L transformer decoder (288d, 8 heads, partial RoPE, SiLU/GELU)
• Parameters: 27M
• Languages: German (fine-tuned from English moonshine-tiny)
• WER: 36.7% on MLS German test set
• Training data: MLS German (~1,967 hours), 10k steps
• Output: Lowercase only, no punctuation
• License: MIT (inherited from upstream)
• Source: dattazigzag/moonshine-tiny-de

Usage with CrispASR

# Explicit model path
./build/bin/crispasr --backend moonshine -m moonshine-tiny-de-dattazigzag-q4_k.gguf -f audio.wav

Notes

• Moonshine models run on CPU only (GPU not needed for these small models)
• Tokenizer (tokenizer.bin) must be in the same directory as the model file
• For better German quality, consider moonshine-base-de-fidoriel (6.9% WER) or moonshine-tiny-de-fidoriel (11.4% WER)