OPEN LINK ↗
// 1h agoMODEL RELEASE
Underdog Bark 3.0 packs ASR into 163MB
Underdog’s Bark 3.0 is a 0.6B English speech-recognition model with a 163 MB download and reported 4.94% average word error rate. It approaches Parakeet accuracy while enabling fast, private local transcription. [Model card](https://huggingface.co/ConwayResearch/Underdog-Bark-3.0)
// ANALYSIS
Bark 3.0’s biggest achievement is practical compression, making high-quality speech recognition small enough to embed in local apps.
- –Matches NVIDIA Parakeet TDT 0.6B v3’s reported accuracy at roughly one-fifteenth the download size
- –Runs about 41× real time on an Apple M2, making meeting notes and dictation viable without cloud costs
- –Derived from Parakeet, so the core innovation is quantization, packaging, and deployment rather than a new architecture
- –English-only, and the 4.94% benchmark is self-reported rather than independently verified on the leaderboard
- –Developers should test memory use on long recordings and comply with its CC BY 4.0 license
// TAGS
bark-3.0speechsttopen-weightsedge-ailocal-first
DISCOVERED
1h ago
2026-10-25
PUBLISHED
1h ago
2026-10-25
RELEVANCE
9/ 10
AUTHOR
AI Search