ConwayResearch/Underdog-Bark-3.0-Clean

Underdog Bark 3.0 Clean

Underdog Bark 3.0 Clean is the clean version of Bark 3.0 Plus: a 0.6B English speech recognition model in a 180 MB download, with speech cleanup built in. It drops "um" and "uh", filler phrases like "you know" and "I mean", false starts and repeated words, turning spoken thoughts into ready-to-use text without a second cleanup model.

Its foundation, Bark 3.0 Plus, averages 4.76% word error across the Open ASR Leaderboard's seven English sets—better than Parakeet TDT 0.6B v3 (4.94%), Qwen3-ASR-0.6B (5.16%) and Whisper large-v3 (6.04%) in the published comparison, at a fraction of their download sizes. Plus reaches 7.76% on meetings (AMI) and 1.92% on parliamentary speech (VoxPopuli). Clean builds on Plus for dictation and readable transcripts rather than preserving every hesitation.

Clean reaches 4.08% WER against cleaned DisfluencySpeech transcripts, and matches Plus's 1.71% on LibriSpeech clean, with 3.78% on LibriSpeech other. These are Clean's results; the 4.76% seven-set average above belongs to Plus, the verbatim model.

It is the speech model in Underdog, where it turns what you say into text for Speak and for meeting notes, on your Mac or iPhone. Your audio never leaves the device.

The Bark family's accuracy–size comparison for verbatim transcription. Bark 3.0 Plus is the foundation of Clean; this chart shows Plus, Pro and the standard model, not Clean's cleanup scores. See the separate Clean and family benchmark tables below.

Highlights

• The clean version of Bark 3.0 Plus: the same 180 MB download class, with built-in disfluency removal for dictation and notes.
• Built-in cleanup: no "um", no false starts, no second model to clean up after it.
• More accurate than Bark 2.0 on all seven Open ASR Leaderboard English sets (5.83 % against 7.48 %), from a download less than half the size.
• As accurate as the verbatim Bark 3.0 Plus on clean read speech (1.71 % on LibriSpeech clean).
• Small enough for a phone: 180 MB to download, 1.25 GB in memory once unpacked.

Benchmarks

Cleanup

DisfluencySpeech test set (250 clips of unscripted conversation), word error rate (%), lower is better. Fillers count as errors.

Model	Download	Against cleaned transcripts	Against verbatim transcripts
Bark 3.0 Clean	180 MB	4.08	13.85
Bark 2.0	472 MB	3.83	13.01
Bark 3.0 (verbatim)	163 MB	16.53	3.15

A cleanup model should score low in the first column. A verbatim model scores low in the second.

Ordinary speech

Model	Download	LS clean	LS other	AMI	Earnings-22	GigaSpeech	SPGISpeech	VoxPopuli	Average
Bark 3.0 Clean	180 MB	1.71	3.78	12.12	8.55	9.76	2.66	2.23	5.83
Bark 2.0	472 MB	2.94	5.21	16.16	11.08	10.56	3.47	2.90	7.48

Word error rate (%) on the full test sets of the Open ASR Leaderboard's seven English sets, scored with its code. Their references are verbatim, so every filler or repeat that Bark 3.0 Clean removes on purpose counts against it on AMI, GigaSpeech and Earnings-22. For verbatim transcription, use Bark 3.0 (4.94) or Bark 3.0 Plus (4.76).

The Plus foundation: accuracy in a small download

Clean comes from Bark 3.0 Plus. The table below reproduces the family's reported verbatim-recognition results and selected baselines, showing the accuracy and compact size behind the Clean version. Clean's own seven-set result is 5.83%, as shown above; cleanup deliberately changes the output style.

Model	Download	LS clean	LS other	AMI	Earnings-22	GigaSpeech	SPGISpeech	VoxPopuli	Average
Bark 3.0 Pro	1,254 MB	1.38	3.11	7.16	5.65	7.46	3.54	1.70	4.29
Bark 3.0 Plus	180 MB	1.71	3.82	7.76	6.70	7.77	3.64	1.92	4.76
Bark 3.0	163 MB	1.89	4.08	7.82	7.05	7.95	3.72	2.05	4.94
Phonon-2†	164 MB	1.73	3.92	9.32	6.96	8.35	3.70	2.46	5.21
Parakeet TDT 0.6B v3†	2,508 MB	1.52	3.13	9.42	5.85	7.99	3.63	3.05	4.94
Qwen3-ASR-0.6B†	1,876 MB	1.70	4.01	9.33	7.83	7.62	2.74	2.90	5.16
Whisper large-v3†	9,261 MB	1.56	3.52	13.63	7.99	8.35	2.71	4.53	6.04
Whisper large-v3-turbo†	1,618 MB	2.13	3.71	13.88	8.09	8.47	2.79	7.02	6.58

WER (%), lower is better, on the full test sets of the seven English datasets used by the Open ASR Leaderboard, including its cleaned versions of AMI, Earnings-22, GigaSpeech and VoxPopuli. Average is the seven-set mean. † Published leaderboard results as of 9 October 2026; the Bark rows use its scoring code and were reported separately, not submitted leaderboard entries. Download sizes are weight-file sizes, not runtime memory. These figures reproduce the Plus model card's comparison and are not a new evaluation of this Clean release.

In Underdog

Underdog runs this file as it is, with no Python. A native Swift runtime reads model.bark, unpacks it in memory in about 2 seconds on Apple silicon, and runs the model on MLX. The unpacked weights are bit-identical to what bark_loader.py produces. Underdog downloads two files, model.bark and mlx/config.json, and checks each against a pinned size and SHA-256 before using it.

On an M5 Max, a 136-second recording transcribes in about 4 seconds, model load included.

Run it yourself

pip install "transformers>=5.19" torch constriction numpy soundfile soxr
python bark_loader.py model.bark recording.wav

from bark_loader import load_model, transcribe
import soundfile as sf

model = load_model("model.bark", device="cuda")          # or "cpu", "mps"
wave, sr = sf.read("recording.wav", dtype="float32", always_2d=True)
print(transcribe(model, wave.mean(1), sr))

On Apple silicon with MLX:

pip install mlx mlx-audio constriction numpy soundfile soxr
python bark_mlx.py recording.wav

The first run unpacks the weights into an MLX folder and caches it. Unpacking is exact: the weights you load are the weights that were measured. Bark 3.0 Clean was measured with no silence added before the audio, which is what transcribe does.

Files

File	What it is
model.bark	The weights, entropy coded (format bark-levels-v3)
mlx/config.json	Model configuration and vocabulary
mlx/tokenizer.*, mlx/vocab.txt	Tokenizer files for mlx-audio
bark_loader.py	Unpacks model.bark into a transformers ParakeetForTDT
bark_mlx.py	Unpacks model.bark into an mlx-audio Parakeet folder

Notes

Based on parakeet-tdt-0.6b-v3 by NVIDIA, through moondream/parakeet-ultra and Bark 3.0 Plus; the tokenizer and output conventions (punctuation, casing, numerals) are the original's. Licence CC-BY-4.0, same as the original; NOTICE lists the attribution. English only.