Parakeet CTC 1.1B

May 11, 2026 · View on GitHub

NVIDIA's nvidia/parakeet-ctc-1.1b ported to transcribe.cpp. A 1.1B-parameter FastConformer-XL encoder with a linear CTC head.

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.

This is the largest pure-CTC variant in the parakeet family and trades size for accuracy. NVIDIA reports a 0.04 percentage-point improvement on LibriSpeech test-clean over the 0.6B sibling.

See NVIDIA's model card for training data, intended use, and upstream evaluation methodology.

Licensed CC-BY-4.0. Ported from upstream commit a707e81, pinned 2026-05-10.

Download

QuantizationDownloadSizeWER (LibriSpeech test-clean)
F32parakeet-ctc-1.1b-F32.gguf4.25 GB1.85%
F16parakeet-ctc-1.1b-F16.gguf2.13 GB1.85%
Q8_0parakeet-ctc-1.1b-Q8_0.gguf1.26 GB1.85%
Q6_Kparakeet-ctc-1.1b-Q6_K.gguf1.04 GB1.85%
Q5_K_Mparakeet-ctc-1.1b-Q5_K_M.gguf929 MB1.84%
Q4_K_Mparakeet-ctc-1.1b-Q4_K_M.gguf818 MB1.90%

WER is measured on the full LibriSpeech test-clean split (2620 utterances) with greedy CTC decoding and no external LM. F32 reference baseline: 1.85%. NVIDIA's self-reported number on the same split is 1.83% (from the HF model card).

Quick Start

cmake -B build
cmake --build build

build/bin/transcribe-cli \
  -m models/parakeet-ctc-1.1b/parakeet-ctc-1.1b-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). Cells gated on Tctl < 55°C per backend.

Apple M4 Max

BackendSampleQ8_0Q4_K_M
Metaljfk (11.0s)91 ms (121×)93 ms (118×)
Metaldots (35.3s)224 ms (158×)224 ms (158×)
CPUjfk (11.0s)602 ms (18×)501 ms (22×)
CPUdots (35.3s)2.04 s (17×)1.70 s (21×)

macOS 26.4.1, transcribe.cpp a6c097e.

AMD Ryzen 7 4750U Pro

BackendSampleQ8_0Q4_K_M
Vulkanjfk (11.0s)825 ms (13×)822 ms (13×)
Vulkandots (35.3s)2.34 s (15×)2.33 s (15×)
CPUjfk (11.0s)1.75 s (6×)1.38 s (8×)
CPUdots (35.3s)6.08 s (6×)5.12 s (7×)

Fedora 43, transcribe.cpp 57997dc. Vulkan device: AMD Radeon Graphics (RADV RENOIR).

Benchmark reproduction:

uv run scripts/bench/run.py \
  --models parakeet-ctc-1.1b \
  --quants q8_0,q4_k_m \
  --samples jfk,dots \
  --backends metal,cpu,vulkan \
  --iters 3 --warmup 1 \
  --name parakeet-ctc-1.1b-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 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).

FieldValue
ReferenceNeMo, nvidia/parakeet-ctc-1.1b
Dump scriptscripts/dump_reference_parakeet_nemo.py
Manifesttests/golden/parakeet/parakeet-ctc-1.1b.manifest.json
Commanduv run scripts/validate.py all --family parakeet --variant parakeet-ctc-1.1b

Reproduction

Convert

uv run --project scripts/envs/parakeet \
  scripts/convert-parakeet.py nvidia/parakeet-ctc-1.1b

Quantize

build/bin/transcribe-quantize \
  models/parakeet-ctc-1.1b/parakeet-ctc-1.1b-F32.gguf \
  models/parakeet-ctc-1.1b/parakeet-ctc-1.1b-Q8_0.gguf \
  --quant Q8_0

Validate

uv run scripts/validate.py all --family parakeet --variant parakeet-ctc-1.1b