cstr/Hojo-ASR-Multi-V1-GGUF

Hojo-ASR-Multi-V1 — GGUF

GGUF conversions of HojoAI/Hojo-ASR-Multi-V1 for CrispASR (--backend hojo-asr).

Hojo-ASR-Multi-V1 is an encoder–adapter–LLM ASR model for German, French, Italian, Portuguese and Spanish:

component	details
audio encoder	Qwen3-Omni audio tower, 32 layers, d=1280, conv stem fusing 8 mel frames into one 1280-d frame
adapter	2 WeNet conformer blocks, d=2560, 1:1 in time (no frame stacking)
decoder	Qwen3-4B-Instruct-2507, lm_head tied to the embeddings
front end	whisper-large-v3 feature extractor (128 mel bins)

All weights come from the checkpoint's single merged_full_model.safetensors. The 30B Qwen3-Omni checkpoint named in its config is only read for audio hyper-parameters and is never downloaded.

Files

File	Size
hojo-asr-multi-v1-f16.gguf	9.6 GB
hojo-asr-multi-v1-q4_k.gguf	4.4 GB (default for -m hojo-asr)

Verification against the upstream package

A reference dump comes from the upstream Python package, via forward hooks, on German and French FLEURS-style clips. CrispASR's crispasr-diff compares every stage:

stage (F16, worst frame cos)	German	French
mel spectrogram	1.000000	1.000000
encoder output	0.999998	0.999999
adapter output	0.999994	0.999995
prefill embeddings	0.999997	0.999997
first-step logits	1.000000	1.000000

End to end, the CrispASR CLI's greedy transcript is identical to the upstream package's greedy transcript on both clips at F16, and at Q4_K on French. At Q4_K on German, one word differs ("weit" for "breed").

Decoding is greedy by default. The model's config.yaml names beam 4, available with -bs 4, but CrispASR's beam search replays each beam's prefix (O(B·T²)) and its ranking does not yet reproduce transformers' generate(). Upstream's own greedy output is what the CLI matches.

Usage

crispasr -m hojo-asr -f audio.wav            # auto-download Q4_K
crispasr --backend hojo-asr -m hojo-asr-multi-v1-f16.gguf -f audio.wav -l de

Licence

Apache-2.0, as the upstream model. Conversion and runtime: CrispASR.