Underdog releases Bark 3.0, a tiny open speech model that runs on phones
Underdog AI released Bark 3.0, a speech recognition model it says is just a 163 MB download and runs on both laptops and phones. The company claims it transcribes a minute of audio in under a quarter second on a MacBook and, on recorded meetings in the AMI benchmark, gets 7.82% of words wrong versus 9.42% for NVIDIA’s Parakeet and 13.63% for OpenAI’s Whisper large-v3. Bark is free and open source under CC BY 4.0, which makes private, on-device transcription easier for anyone. These are Underdog’s own claims.