cstr/quds-v4-fa-GGUF

Quds v4 (Persian ASR) — GGUF for CrispASR

GGUF conversion of hojreh/Quds-v4-onnx (Quds v4, by @hojreh) — a Persian ASR model fine-tuned from nvidia/stt_fa_fastconformer_hybrid_large (FastConformer Hybrid, CC-BY-4.0). Architecture: FastConformer encoder (17 layers, d=512) + single-LSTM RNNT prediction network; trained on ~600 h of Persian audio, optimized for Islamic lectures and Howzah courses. The RNNT decoder is included; the original hybrid's CTC head is not part of the upstream ONNX export.

License: CC-BY-NC-4.0 — non-commercial use only (upstream model license). Attribution: Quds v4 by hojreh; base model by NVIDIA (CC-BY-4.0). The mel filterbank/window buffers are copied from the NVIDIA base checkpoint.

file	quant	size	notes
quds-v4-fa-q8_0.gguf	Q8_0	122 MB	recommended — transcripts byte-identical to F16 on our test set
quds-v4-fa-q4_k.gguf	Q4_K	77 MB	small; minor word-level drift vs F16
quds-v4-fa-f16.gguf	F16	219 MB	reference precision

Usage (CrispASR)

crispasr --backend quds -m auto -l fa audio.wav     # auto-download (accepts the NC license)

Validated against the upstream ONNX under onnx_asr (the model card's own runner): transcripts byte-identical on 4/5 real Common Voice Persian clips, with the fifth differing only at the clip onset (where this conversion is closer to the human reference). Converted with CrispASR's models/convert-nemo-rnnt-onnx-to-gguf.py (#387).