Hugging Face's logo

UsefulSensors
/
moonshine-base

Automatic Speech Recognition
Transformers
Safetensors
English
moonshine
Eval Results
Model card Files
xet
 Community
6

GGUF + pure-C++ runtime in CrispASR — Moonshine base

#6
by cstr - opened
Discussion
cstr

We've added Moonshine-base to CrispASR — same moonshine backend as tiny, just dispatched on GGUF metadata (moonshine-impl.h shared between sizes).

Runtime details:

  • Conv stem + 8L transformer encoder + 8L decoder (416d, partial RoPE, SiLU).
  • KV-cached autoregressive decode with flash attention.
  • Companion-file mechanism for tokenizer.bin on auto-download.

Pre-quantised GGUFs (MIT): cstr/moonshine-base-GGUF — plus the per-language variants we converted: cstr/moonshine-base-{ja,ko,zh,ar,vi,uk}-GGUF.

./build/bin/crispasr --backend moonshine -m moonshine-base-q4_k.gguf -f audio.wav -osrt

Companion tiny size: cstr/moonshine-tiny-GGUF. Streaming variants: cstr/moonshine-streaming-{tiny,small,medium}-GGUF (separate moonshine-streaming backend).

Sign up or log in to comment