Skip to content

parakeet-ctc-1.1b.md

Latest commit

 

History

History
134 lines (100 loc) · 5.21 KB

parakeet-ctc-1.1b.md

File metadata and controls

134 lines (100 loc) · 5.21 KB
Raw

Parakeet CTC 1.1B

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

Quantization Download Size WER (LibriSpeech test-clean)
F32 parakeet-ctc-1.1b-F32.gguf 4.25 GB 1.85%
F16 parakeet-ctc-1.1b-F16.gguf 2.13 GB 1.85%
Q8_0 parakeet-ctc-1.1b-Q8_0.gguf 1.26 GB 1.85%
Q6_K parakeet-ctc-1.1b-Q6_K.gguf 1.04 GB 1.85%
Q5_K_M parakeet-ctc-1.1b-Q5_K_M.gguf 929 MB 1.84%
Q4_K_M parakeet-ctc-1.1b-Q4_K_M.gguf 818 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: 1.85%. 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

Backend Sample Q8_0 Q4_K_M
Metal jfk (11.0s) 91 ms (121×) 93 ms (118×)
Metal dots (35.3s) 224 ms (158×) 224 ms (158×)
CPU jfk (11.0s) 602 ms (18×) 501 ms (22×)
CPU dots (35.3s) 2.04 s (17×) 1.70 s (21×)

macOS 26.4.1, transcribe.cpp a6c097e.

AMD Ryzen 7 4750U Pro

Backend Sample Q8_0 Q4_K_M
Vulkan jfk (11.0s) 825 ms (13×) 822 ms (13×)
Vulkan dots (35.3s) 2.34 s (15×) 2.33 s (15×)
CPU jfk (11.0s) 1.75 s (6×) 1.38 s (8×)
CPU dots (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).

Field Value
Reference NeMo, nvidia/parakeet-ctc-1.1b
Dump script scripts/dump_reference_parakeet_nemo.py
Manifest tests/golden/parakeet/parakeet-ctc-1.1b.manifest.json
Command uv 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