scholzmx/underdog-bark-3.0-plus-gguf

Underdog Bark 3.0 Plus — phone GGUF (Starling engine)

A Starling GGUF of ConwayResearch/Underdog-Bark-3.0-Plus by Conway Research, the 180 MB verbatim member of the family. It is a fine-tune of nvidia/parakeet-tdt-0.6b-v3 (FastConformer encoder + TDT transducer, English), so it runs on Starling's Parakeet engine unchanged (--model parakeet), including the Vulkan fast engine in the Android app.

Runtime note: these files follow the Starling GGUF tensor contract and run on the native starling-serve binary / libstarling_ggml engine from the starling repository, including the Android app. They are not llama.cpp / whisper.cpp GGUFs.

Output style (inherited from the source model, reproduced exactly): on roughly 1 in 10 FLEURS clips this model writes lowercase text without punctuation, and it often spells numbers out ("twenty five" instead of "25"). Bark 3.0 Pro does not do this.

File

file	size	layout
underdog-bark-3.0-plus-q4pw.gguf	364 MB	every linear Q4_0 (incl. the conformer pointwise convs); joint, LSTM and prediction embedding Q8_0 — the same layout as Starling's phone Parakeet

SHA-256: 586e2d0306a82cf87c6d47e1fff7b80484a4143eaf5afb7da9b4a87cbcab3d29

Measured quality

FLEURS en_us test, 100 clips, every file decoded in its own process on the ggml CPU engine. "Exact" is the lossless conversion of the published weights (its transcripts match the transformers reference word for word on 60 clips). Two scorings: Starling's own text normalization, and Whisper's English normalizer as used by the Open ASR Leaderboard (it also maps spelled-out numbers to digits, which matters for this family).

	exact	q4pw	paired Δ (q4pw − exact), 95 % CI
WER, Starling normalization	7.17 %	7.12 %	−0.08 [−0.40, +0.17] pt
WER, Whisper normalization	4.52 %	4.52 %
stock Parakeet TDT 0.6B v3 (bf16), same clips	5.56 %		Whisper normalization: 4.29 %

FLEURS is read Wikipedia text; it does not test the conversational speech (meetings, earnings calls) the publisher's Open ASR numbers cover.

How it was built

1. Convert the published weights exactly (model.bark): uv run --group convert python scripts/convert_parakeet_gguf.py --bark model.bark --dtype f16 --output exact.gguf
2. Importance matrix: 100 FLEURS en_us validation clips through the exact file (STARLING_IMATRIX=…, benchmarks/quant_phone/wer_models.py); included here.
3. starling-quantize --input exact.gguf --output underdog-bark-3.0-plus-q4pw.gguf --recipe parakeet-q4pw-fullimx.recipe --imatrix underdog-bark-3.0-plus.imx --shrink-f16

Usage

hf download scholzmx/underdog-bark-3.0-plus-gguf underdog-bark-3.0-plus-q4pw.gguf --local-dir ./models
starling-serve --model parakeet --gguf ./models/underdog-bark-3.0-plus-q4pw.gguf --port 8181

License and attribution

CC-BY-4.0, as the source. Underdog Bark 3.0 Plus is by Conway Research (ConwayResearch/Underdog-Bark-3.0-Plus); it derives from nvidia/parakeet-tdt-0.6b-v3 (NVIDIA) and moondream/parakeet-ultra (Moondream). Their NOTICE is included unchanged. Changes in this repository: format conversion to GGUF and requantization to the Q4_0/Q8_0 layout above.