Underdog drops Bark 3.0: A 163MB speech model that beats Whisper Large

Underdog just released Bark 3.0, a remarkably compact, local speech-to-text model designed to run directly on consumer hardware without breaking a sweat.

The base model download is just 163 MB. On an M5 Max MacBook Pro, it can chew through a five-minute audio recording in roughly half a second, which is about 595 times faster than real-time playback. Better yet, its accuracy in conversational English beats out established alternatives that are dozens of times larger.

Shrinking down without losing precision

Bark 3.0 is built on top of NVIDIA's parakeet-tdt-0.6b-v3 and Moondream's parakeet-ultra architectures, both distributed under permissive CC-BY 4.0 terms.

The family includes four distinct weights. The standard Bark 3.0 sits at 163 MB. If you want slightly fewer mistakes, there is a 180 MB Plus version. For pure precision, there is a 1.25 GB Pro variant that easily fits inside virtually any consumer GPU VRAM. Finally, there is an 180 MB Clean version built specifically for spoken dictation.

On the Open ASR Leaderboard across seven standard English benchmarks, the 163 MB Bark 3.0 posted an average word error rate of 4.94%. The 180 MB Plus version improved that to 4.76%, while the 1.25 GB Pro reached 4.29%.

To put those numbers in perspective, OpenAI's Whisper large-v3 weighs over 9.2 GB and averaged a 6.04% error rate. NVIDIA's 2.5 GB Parakeet hit 4.94%. Bark matches or beats both while being a fraction of their footprint.

The real test comes in noisy meetings. On the AMI meeting benchmark, Whisper large-v3 misses 13.63% of words, and Parakeet misses 9.42%. The 163 MB Bark 3.0 drops that error rate down to 7.82%.

Filtering the fluff with Bark Clean

Real conversations are full of false starts, stutters, and verbal pauses. If you transcribe every sound verbatim, meeting summaries and dictation turn into a chaotic mess.

Bark 3.0 Clean is trained explicitly to output what you meant rather than every single noise. Stumbles like repeated phrases, filler expressions, and hesitations are discarded automatically. An audio clip containing false starts and filler words produces a crisp, ready-to-read sentence.

That said, there are several caveats you should keep in mind before jumping in. First, it is strictly English only. Second, the model does not yet output word-level timestamps or distinguish between different speakers in a room. Third, user-defined custom vocabulary lists supported in earlier releases do not work with 3.0 yet.

How to get started

All weights are hosted directly on Hugging Face under terms that permit commercial usage with attribution. You can try a live web demo directly in your browser by clicking record, or grab the weights and follow the local setup instructions on the model card.

Why it matters

Bark 3.0 makes a compelling case for specialized, compact local intelligence. For years, running high-quality speech recognition meant dealing with multi-gigabyte models or pinging remote cloud APIs. Shrinking that capability down to 163 MB while maintaining sub-second latency completely shifts what is practical on ordinary laptops.

If you build local transcription software, voice dictation tools, or desktop productivity apps that focus on English speech, this model is an immediate upgrade over Whisper Large. It saves system resources, protects user privacy entirely on-device, and delivers a snappier user experience without sacrificing accuracy.