Parakeet TDT 0.6B v2 is NVIDIA's English-only 600 M-parameter ASR model — the original Open ASR Leaderboard topper before v3 spread capacity across 25 European languages. On plain English, v2 is often stronger than v3 since it didn't have to share encoder capacity with 24 other languages.

English-only, mixed-case + punctuation output
Built-in word-level timestamps from the TDT (Token-and-Duration Transducer) decoder — no separate CTC alignment model required
CC-BY-4.0 licence (friendlier than most ASR models)

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-0.6b-v2.gguf`	1.24 GB	F16, full precision
`parakeet-tdt-0.6b-v2-q8_0.gguf`	735 MB	Q8_0, near-lossless
`parakeet-tdt-0.6b-v2-q4_k.gguf`	468 MB	Q4_K — recommended default

All three precisions produce the same text on samples/jfk.wav:

And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.

Quick Start

bash

# 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-v2 --auto-download -f your-audio.wav

# 2b. Or explicit download + load
hf download cstr/parakeet-tdt-0.6b-v2-GGUF \
    parakeet-tdt-0.6b-v2-q4_k.gguf --local-dir .
./build/bin/crispasr -m parakeet-tdt-0.6b-v2-q4_k.gguf -f your-audio.wav

When to pick v2 over v3

Scenario	Pick
English only, want best WER	v2 (this repo)
Multilingual, 25 EU languages	v3 — `cstr/parakeet-tdt-0.6b-v3-GGUF`
Tight RAM, English	smaller hybrid — `cstr/parakeet-tdt_ctc-110m-GGUF`
Long-tail English vocab, willing to pay 2x compute	larger — `cstr/parakeet-tdt-1.1b-GGUF`

Model architecture

Component	Details
Encoder	24-layer FastConformer, d=1024, 8 heads, head_dim=128, FFN=4096, conv kernel=9
Subsampling	Conv2d dw_striding stack, 8× temporal (50 → 12.5 fps)
Predictor	2-layer LSTM, hidden 640
Joint head	enc(1024 → 640) + pred(640 → 640) → ReLU → linear(640 → 1029)
Vocab	1024 SentencePiece tokens (English, mixed case + punctuation)
Audio	16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400
Parameters	~600 M

Same FastConformer encoder + TDT decoder as v3 — just trained on English-only data with an English-only BPE.

How this was made

The .nemo checkpoint was unpacked, NeMo state-dict keys were remapped to ggml-friendly names, and weights were written to GGUF F16 (matmul tensors) + F32 (norms / biases / mel filterbank).
Quantised variants are produced by crispasr-quantize (the same llama.cpp-style quantiser used for the other GGUF releases).
Inference uses src/parakeet.{h,cpp}: FastConformer encoder runs as a single ggml graph (BN folded out), LSTM predictor + joint head run as CPU F32 loops, TDT greedy decode alternates "advance encoder frame" / "emit token + advance predictor" using the duration head's argmax.

Attribution

Original model: nvidia/parakeet-tdt-0.6b-v2 (CC-BY-4.0). NVIDIA NeMo team.
GGUF conversion + ggml runtime: CrispStrobe/CrispASR.

License

CC-BY-4.0, inherited from the base model. Use of these GGUF files must comply with the CC-BY-4.0 license including attribution.