cstr/parakeet-redux-GGUF

parakeet-redux — GGUF (CrispASR)

GGUF conversion of moondream/parakeet-redux for CrispASR. Use -m parakeet-redux to auto-download Q4_K, or --backend parakeet -m <gguf>.

parakeet-tdt-0.6b-v3 architecture with a ternary encoder: weights in {-1, 0, +1} × a per-group scale, stored base-3 packed (5 digits per byte). convert-parakeet-to-gguf.py --hf unpacks them exactly (w = scale · (code − 1)). The GGUF therefore holds ordinary F16/Q80/Q4K weights and runs on the unmodified parakeet runtime.

Verification

Reference: transformers ParakeetForTDT. The independent check is moondream's own Photon runtime.

• —Diff harness (F16): mel, subsampling, all 24 encoder layers and the encoder output pass, on jfk (11 s, English) and a German clip.
• —Transcripts: F16, Q80 and Q4K are identical to transformers and to Photon on both clips.

en: And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
de: Guten Morgen, die Sitzung beginnt heute um neun Uhr im großen Saal.

Licence: CC-BY-4.0, as the original.