stt-kk-ru-fastconformer-hybrid-ctc-large-GGUF

GGUF conversions of the CTC branch of nvidia/stt_kk_ru_fastconformer_hybrid_large for CrispASR. The upstream model is a hybrid transducer+CTC Kazakh + Russian ASR release; the shared FastConformer encoder plus the auxiliary CTC head are extracted here as a standalone CTC model (the RNNT prediction network and joint are dropped), giving a compact Kazakh + Russian ASR and forced-alignment model (no punctuation/casing).

Quant	Size	Description
F16	219 MB	Full precision
Q8_0	130 MB	8-bit
Q4_K	82 MB	4-bit K-quant (recommended)

Architecture

17-layer NeMo FastConformer encoder + Conv1d CTC head. d_model=512, 8 heads, SentencePiece vocab, 80 log-mel features, ~115M params.

Usage

# Kazakh + Russian ASR:
crispasr --backend fastconformer-ctc -m stt-kk-ru-fastconformer-hybrid-ctc-large-q4_k.gguf -f audio.wav

# Forced alignment (word timestamps for known text, or re-timing an .srt):
crispasr --align-only -am stt-kk-ru-fastconformer-hybrid-ctc-large-q4_k.gguf \
    -f audio.wav --text-file subtitles.srt --align-output retimed.srt

Attribution

All credit for the model goes to NVIDIA's NeMo team; this repository only repackages the CTC branch in GGUF form under the same CC-BY-4.0 license. Conversion: models/convert-stt-fastconformer-ctc-to-gguf.py in CrispASR.

Provenance and EU AI Act Art. 53 note

• Upstream model: nvidia/stt_kk_ru_fastconformer_hybrid_large — published by nvidia.
• Upstream licence: cc-by-4.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.