Moonshine Tiny (German, fidoriel) -- GGUF

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

Available variants

File Quant Size Notes
moonshine-tiny-de-fidoriel-f16.gguf F16 52 MB Half precision
moonshine-tiny-de-fidoriel-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: 11.4% on Common Voice 22 German test set
  • CER: 4.2%
  • Output: Proper casing and punctuation
  • License: CC-BY-NC-SA-4.0 (inherited from upstream)
  • Source: fidoriel/moonshine-tiny-de

Usage with CrispASR

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

# Or via backend name (auto-download)
./build/bin/crispasr --backend moonshine-tiny-de -m auto -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
  • Smaller/faster alternative to moonshine-base-de (17 MB vs 39 MB)
Downloads last month
133
GGUF
Model size
27.1M params
Architecture
moonshine
Hardware compatibility
Log In to add your hardware

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for cstr/moonshine-tiny-de-fidoriel-GGUF

Quantized
(1)
this model