Parakeet Tdt Ctc 110m GGUF

• English-only, mixed-case + punctuation output
• Hybrid TDT+CTC head — runtime defaults to CTC decode (the TDT path needs a 2-LSTM predictor which this model doesn't have; single-LSTM predictor + CTC head is a deliberate trade for size)
• ~45× realtime on M1 Metal with Q4_K — fastest in the family
• CC-BY-4.0 licence

This repo provides three quantisations, all converted from the same .nemo checkpoint via the convert-parakeet-to-gguf.py script and quantised with crispasr-quantize.

Files

Smoke test on samples/jfk.wav (11 s clip, M1 Metal):

Note: the model emits "askk" instead of "ask" on this clip — a small-model artifact, not a runtime bug. Larger Parakeet variants (v2 / tdt-1.1b / tdt_ctc-1.1b) get it right.

Quick Start

# 1. Build the runtime
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target crispasr

# 2a. Auto-download via the registry key
./build/bin/crispasr -m parakeet-tdt_ctc-110m --auto-download -f your-audio.wav

# 2b. Or explicit download + load
hf download cstr/parakeet-tdt_ctc-110m-GGUF \
    parakeet-tdt_ctc-110m-q4_k.gguf --local-dir .
./build/bin/crispasr -m parakeet-tdt_ctc-110m-q4_k.gguf -f your-audio.wav

The runtime detects pred_layers=1 + has_ctc=True at load time and automatically flips to CTC decode — no flag needed. See parakeet_init_from_file in src/parakeet.cpp.

When to pick this over the other Parakeet variants

Model architecture

Why the runtime auto-flips to CTC

Hybrid TDT+CTC models normally let the user choose between the two decoders. But this checkpoint ships with pred_layers=1 — a single LSTM rather than the usual two — which is too shallow for the TDT prediction network to produce useful joint scores. Upstream's recommended decode is CTC for this checkpoint. The C++ runtime detects this at load (pred_layers < 2 && has_ctc) and sets decode_ctc=true automatically. You can opt back into TDT (not recommended) by removing that auto-flip in parakeet_init_from_file.

Attribution

• Original model: nvidia/parakeet-tdt_ctc-110m (CC-BY-4.0). NVIDIA NeMo team.
• GGUF conversion + ggml runtime: CrispStrobe/CrispASR.

License

CC-BY-4.0, inherited from the base model.

Provenance and EU AI Act Art. 53 note

• Upstream model: nvidia/parakeet-tdt_ctc-110m — 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/GGML). 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.

parakeet-tdt_ctc-110m.gguf

gguf · 219 MB · SHA-256 ec7e1702cee4…92f4 · Hugging Face