cstr/parakeet-tdt-1.1b-GGUF

Parakeet TDT 1.1B — GGUF (ggml-quantised)

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

The larger Parakeet TDT — 1.1 B parameters, 42-layer FastConformer encoder. The biggest pure-English TDT variant in the family. Pick this when you want maximum WER quality on long-tail English vocabulary and don't mind paying 2× the compute relative to 0.6 B.

• English-only, lowercase output without punctuation
• Built-in word-level timestamps from the TDT decoder
• 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-1.1b.gguf	2.14 GB	F16, full precision
parakeet-tdt-1.1b-q8_0.gguf	1.27 GB	Q8_0, near-lossless
parakeet-tdt-1.1b-q4_k.gguf	808 MB	Q4_K — recommended default

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

Quant	Time	Realtime	Output
F16	0.68 s	16.1×	"and so my fellow americans ask not what your country can do for you ask what you can do for your country"
Q8_0	0.67 s	16.4×	(identical)
Q4_K	0.69 s	16.0×	(identical)

Output is lowercase, no punctuation by design — the upstream vocab is lowercase-only. If you need proper casing/punctuation, pipe the output through a punctuation-restoration post-processor (--punc-model fullstop-punc or fireredpunc).

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

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

# 2c. Lowercase output → add punctuation
./build/bin/crispasr -m parakeet-tdt-1.1b --punc-model fullstop-punc -f your-audio.wav

When to pick this over the other Parakeet variants

Scenario	Pick
English, long-tail vocab, fine with 2× compute	tdt-1.1b (this repo)
English, best WER per FLOP, mixed-case output	cstr/parakeet-tdt-0.6b-v2-GGUF
Multilingual (25 EU languages)	cstr/parakeet-tdt-0.6b-v3-GGUF
Tight RAM, English	cstr/parakeet-tdt_ctc-110m-GGUF
English 1.1B with proper casing/punct in output	cstr/parakeet-tdt_ctc-1.1b-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
Vocab	1024 SentencePiece tokens (English, lowercase) + blank
Audio	16 kHz mono, 80 mel bins, n_fft=512, hop=160, win=400
Parameters	~1.1 B

42 layers vs 24 for 0.6b — same encoder design, just deeper.

Attribution

• Original model: nvidia/parakeet-tdt-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.