Parakeet CTC 0.6B
May 11, 2026 · View on GitHub
NVIDIA's nvidia/parakeet-ctc-0.6b
ported to transcribe.cpp. A 0.6B-parameter FastConformer-Large encoder with a
linear CTC head — the simplest decoder in the parakeet family and therefore
the fastest.
What it's for
Offline English speech-to-text with greedy CTC decoding. Output is lowercase, no punctuation (the upstream model card explicitly notes "lower case English alphabet"). Token- and word-level timestamps are available. Not a streaming model; does not translate.
The encoder is identical in shape to parakeet-tdt-0.6b-v2 (24 layers,
1024-d), so this variant is the fastest 0.6B-class option in the family
on this codebase.
See NVIDIA's model card for training data, intended use, and upstream evaluation methodology.
Licensed CC-BY-4.0. Ported from upstream commit
ad09ba1,
pinned 2026-05-10.
Download
| Quantization | Download | Size | WER (LibriSpeech test-clean) |
|---|---|---|---|
| F32 | parakeet-ctc-0.6b-F32.gguf | 2.44 GB | 1.87% |
| F16 | parakeet-ctc-0.6b-F16.gguf | 1.22 GB | 1.87% |
| Q8_0 | parakeet-ctc-0.6b-Q8_0.gguf | 722 MB | 1.87% |
| Q6_K | parakeet-ctc-0.6b-Q6_K.gguf | 594 MB | 1.84% |
| Q5_K_M | parakeet-ctc-0.6b-Q5_K_M.gguf | 533 MB | 1.87% |
| Q4_K_M | parakeet-ctc-0.6b-Q4_K_M.gguf | 469 MB | 1.90% |
WER is measured on the full LibriSpeech test-clean split (2620 utterances) with greedy CTC decoding and no external LM. F32 reference baseline: 1.87%. NVIDIA's self-reported number on the same split is 1.87% (from the HF model card).
Quick Start
cmake -B build
cmake --build build
build/bin/transcribe-cli \
-m models/parakeet-ctc-0.6b/parakeet-ctc-0.6b-Q8_0.gguf \
samples/jfk.wav
If your audio is not already 16 kHz mono WAV, convert it first:
ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
Performance
Cells are wall-clock latency (mean over 3 iterations after 1 warmup),
with speedup over realtime in parentheses. Units: ms below 1 s, s
above (2 decimal places).
Apple M4 Max
| Backend | Sample | Q8_0 | Q4_K_M |
|---|---|---|---|
| Metal | jfk (11.0s) | 58 ms (191×) | 59 ms (185×) |
| Metal | dots (35.3s) | 143 ms (246×) | 145 ms (244×) |
| CPU | jfk (11.0s) | 356 ms (31×) | 298 ms (37×) |
| CPU | dots (35.3s) | 1.19 s (30×) | 1.00 s (35×) |
macOS 26.4.1, transcribe.cpp a6c097e.
AMD Ryzen 7 4750U Pro
| Backend | Sample | Q8_0 | Q4_K_M |
|---|---|---|---|
| Vulkan | jfk (11.0s) | 520 ms (21×) | 537 ms (20×) |
| Vulkan | dots (35.3s) | 1.50 s (24×) | 1.50 s (24×) |
| CPU | jfk (11.0s) | 1.07 s (10×) | 863 ms (13×) |
| CPU | dots (35.3s) | 3.67 s (10×) | 3.14 s (11×) |
Fedora 43, transcribe.cpp 57997dc. Vulkan device: AMD Radeon Graphics (RADV RENOIR).
Benchmark reproduction:
uv run scripts/bench/run.py \
--models parakeet-ctc-0.6b \
--quants q8_0,q4_k_m \
--samples jfk,dots \
--backends metal,cpu,vulkan \
--iters 3 --warmup 1 \
--name parakeet-ctc-0.6b-publication
Numerical Validation
transcribe.cpp is validated tensor-by-tensor against NeMo on samples/jfk.wav
via scripts/validate.py, sharing the parakeet family tolerance file. The
encoder shape is identical to parakeet-tdt-0.6b-v2 and uses the same
FastConformer code path; the family-level forward map at
reports/porting/parakeet/forward-map.md
documents the per-stage divergence sources (fp64 STFT, mel amplification,
attenuation through the encoder).
| Field | Value |
|---|---|
| Reference | NeMo, nvidia/parakeet-ctc-0.6b |
| Dump script | scripts/dump_reference_parakeet_nemo.py |
| Manifest | tests/golden/parakeet/parakeet-ctc-0.6b.manifest.json |
| Command | uv run scripts/validate.py all --family parakeet --variant parakeet-ctc-0.6b |
Reproduction
Convert
uv run --project scripts/envs/parakeet \
scripts/convert-parakeet.py nvidia/parakeet-ctc-0.6b
Quantize
build/bin/transcribe-quantize \
models/parakeet-ctc-0.6b/parakeet-ctc-0.6b-F32.gguf \
models/parakeet-ctc-0.6b/parakeet-ctc-0.6b-Q8_0.gguf \
--quant Q8_0
Validate
uv run scripts/validate.py all --family parakeet --variant parakeet-ctc-0.6b