cstr/parakeet-tdt_ctc-110m-GGUF

Parakeet TDT+CTC 110M — GGUF (ggml-quantised)

GGUF / ggml conversions of nvidia/parakeet-tdt_ctc-110m for use with the crispasr CLI from CrispStrobe/CrispASR.

The smallest Parakeet variant: 110 M parameters, 17-layer FastConformer encoder with both TDT and CTC heads. Designed for low-RAM hosts and high-throughput batch transcription.

• 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

File	Size	Notes
parakeet-tdt_ctc-110m.gguf	230 MB	F16, full precision
parakeet-tdt_ctc-110m-q8_0.gguf	139 MB	Q8_0, near-lossless
parakeet-tdt_ctc-110m-q4_k.gguf	91 MB	Q4_K — recommended default

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

Quant	Time	Realtime	Output
F16	0.24 s	45.7×	"And so, my fellow Americans, askk not what your country can do for you. Ask what you can do for your country."
Q8_0	0.25 s	44.6×	(identical)
Q4_K	0.25 s	43.8×	(identical)

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

Scenario	Pick
English, tightest RAM (mobile / edge / embedded)	110m (this repo)
English, best WER, ~600 M params	cstr/parakeet-tdt-0.6b-v2-GGUF
Multilingual (25 EU languages)	cstr/parakeet-tdt-0.6b-v3-GGUF
English, long-tail vocab	cstr/parakeet-tdt-1.1b-GGUF

Model architecture

Component	Details
Encoder	17-layer FastConformer, d=512, 8 heads, head_dim=64, FFN=2048, conv kernel=9
Subsampling	Conv2d dw_striding stack, 8× temporal (100 → 12.5 fps)
Predictor	1-layer LSTM, hidden 640 (smaller than the standard 2-LSTM)
Joint head	enc(512 → 640) + pred(640 → 640) → ReLU → linear(640 → 1029) — TDT, 5 durations
CTC head	linear(512 → 1025) — used by default at runtime
Vocab	1024 SentencePiece tokens (English) + blank
Audio	16 kHz mono, 80 mel bins, n_fft=512, hop=160, win=400
Parameters	~110 M

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.