parakeet-tdt-1.1b.md

Parakeet TDT 1.1B

NVIDIA's nvidia/parakeet-tdt-1.1b ported to transcribe.cpp. A 1.1B-parameter FastConformer-XL encoder with a TDT/RNN-T transducer decoder (predictor + joint with duration head).

What it's for

Offline English speech-to-text with greedy TDT decoding. Output is lowercase, no punctuation (per the upstream model card). Token-, word- and segment-level timestamps are available. Not a streaming model; does not translate.

The 1.1B-class TDT — same family as the published parakeet-tdt-0.6b-v2 but with a larger encoder. TDT decoding (with its duration logits) generally runs faster than plain RNN-T at comparable accuracy because each step can advance more than one frame.

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

Licensed CC-BY-4.0. Ported from upstream commit 53276c6, pinned 2026-05-10.

Download

Quantization	Download	Size	WER (LibriSpeech test-clean)
F32	parakeet-tdt-1.1b-F32.gguf	4.28 GB	1.39%
F16	parakeet-tdt-1.1b-F16.gguf	2.15 GB	1.39%
Q8_0	parakeet-tdt-1.1b-Q8_0.gguf	1.27 GB	1.38%
Q6_K	parakeet-tdt-1.1b-Q6_K.gguf	1.04 GB	1.40%
Q5_K_M	parakeet-tdt-1.1b-Q5_K_M.gguf	936 MB	1.39%
Q4_K_M	parakeet-tdt-1.1b-Q4_K_M.gguf	825 MB	1.42%

WER is measured on the full LibriSpeech test-clean split (2620 utterances) with greedy TDT decoding and no external LM. F32 reference baseline: 1.39%. NVIDIA's self-reported number on the same split is 1.39% (from the HF model card).

Quick Start

cmake -B build
cmake --build build

build/bin/transcribe-cli \
  -m models/parakeet-tdt-1.1b/parakeet-tdt-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)	98 ms (112×)	97 ms (113×)
Metal	dots (35.3s)	254 ms (139×)	255 ms (139×)
CPU	jfk (11.0s)	608 ms (18×)	505 ms (22×)
CPU	dots (35.3s)	2.04 s (17×)	1.73 s (20×)

macOS 26.4.1, transcribe.cpp 12f1076.

AMD Ryzen 7 4750U Pro

Backend	Sample	Q8_0	Q4_K_M
Vulkan	jfk (11.0s)	966 ms (11×)	942 ms (12×)
Vulkan	dots (35.3s)	3.03 s (12×)	3.00 s (12×)
CPU	jfk (11.0s)	1.92 s (6×)	1.49 s (7×)
CPU	dots (35.3s)	6.82 s (5×)	5.94 s (6×)

Fedora 43, transcribe.cpp 12f1076. Vulkan device: AMD Radeon Graphics (RADV RENOIR).

Benchmark reproduction:

uv run scripts/bench/run.py \
  --models parakeet-tdt-1.1b \
  --quants q8_0,q4_k_m \
  --samples jfk,dots \
  --backends metal,cpu,vulkan \
  --iters 3 --warmup 1 \
  --name parakeet-tdt-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-tdt-1.1b
Dump script	scripts/dump_reference_parakeet_nemo.py
Manifest	tests/golden/parakeet/parakeet-tdt-1.1b.manifest.json
Command	uv run scripts/validate.py all --family parakeet --variant parakeet-tdt-1.1b

Reproduction

Convert

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

Quantize

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

Validate

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