cstr/parakeet-tdt_ctc-1.1b-GGUF

Parakeet TDT+CTC 1.1B — GGUF (ggml-quantised)

GGUF / ggml conversions of nvidia/parakeet-tdt_ctc-1.1b for use with the crispasr CLI from CrispStrobe/CrispASR.

The largest hybrid Parakeet — 1.1 B parameters, 42-layer FastConformer encoder with both TDT and CTC heads. The hybrid head gives you two decode strategies on the same encoder: native TDT word timestamps (default), or CTC if you need shallow-fusion biasing.

• English, mixed-case + punctuation output (vocab includes uppercase + punctuation tokens, unlike the pure parakeet-tdt-1.1b)
• Hybrid TDT+CTC — default decode is TDT; pass --parakeet-decoder ctc for the CTC head
• 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-1.1b.gguf	2.15 GB	F16, full precision
parakeet-tdt_ctc-1.1b-q8_0.gguf	1.27 GB	Q8_0, near-lossless
parakeet-tdt_ctc-1.1b-q4_k.gguf	810 MB	Q4_K — recommended default

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

Quant	Time	Realtime	Output
F16	0.74 s	14.8×	"And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country."
Q8_0	2.12 s	5.2×	(identical)
Q4_K	2.67 s	4.1×	(identical)

Note: this checkpoint's Q4_K/Q8_0 run slower than the pure parakeet-tdt-1.1b quants on M1 (CTC + TDT both wired in, plus a per-tensor q4_0 fallback on the joint head). F16 is the fastest precision here.

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-1.1b --auto-download -f your-audio.wav

# 2b. Or explicit download + load
hf download cstr/parakeet-tdt_ctc-1.1b-GGUF \
    parakeet-tdt_ctc-1.1b-q4_k.gguf --local-dir .
./build/bin/crispasr -m parakeet-tdt_ctc-1.1b-q4_k.gguf -f your-audio.wav

# 2c. Switch to the CTC head (e.g. when adding hotword biasing)
./build/bin/crispasr -m parakeet-tdt_ctc-1.1b --parakeet-decoder ctc -f your-audio.wav

When to pick this over the other Parakeet variants

Scenario	Pick
English 1.1B with proper casing + punctuation in output	tdt_ctc-1.1b (this repo)
English 1.1B, lowercase output, faster Q4_K/Q8_0	cstr/parakeet-tdt-1.1b-GGUF
English, best WER per FLOP	cstr/parakeet-tdt-0.6b-v2-GGUF
Multilingual (25 EU languages)	cstr/parakeet-tdt-0.6b-v3-GGUF
Tight RAM	cstr/parakeet-tdt_ctc-110m-GGUF

Model architecture

Component	Details
Encoder	42-layer FastConformer, d=1024, 8 heads, head_dim=128, FFN=4096, conv kernel=9
Subsampling	Conv2d dw_striding stack, 8× temporal (100 → 12.5 fps)
Predictor	2-layer LSTM, hidden 640
Joint head	enc(1024 → 640) + pred(640 → 640) → ReLU → linear(640 → 1029) — TDT, 5 durations
CTC head	linear(1024 → 1025)
Vocab	1024 SentencePiece tokens (English, mixed case + punctuation) + blank
Audio	16 kHz mono, 80 mel bins, n_fft=512, hop=160, win=400
Parameters	~1.1 B

Same 42-layer encoder as parakeet-tdt-1.1b, but with an added CTC head and a mixed-case + punctuated vocab.

Attribution

• Original model: nvidia/parakeet-tdt_ctc-1.1b (CC-BY-4.0). NVIDIA NeMo team.
• GGUF conversion + ggml runtime: CrispStrobe/CrispASR.

License

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