Underdog releases Bark 3.0, a 163 MB speech model that rivals cloud-based systems

Underdog, a new AI lab by Conway Research, has released Bark 3.0, a speech-to-text model compressed into a 163 MB download that matches or exceeds larger commercial systems while running entirely on-device. The model, built on NVIDIA's Parakeet foundation and released under the CC-BY 4.0 license, achieves a 4.94 percent word error rate across the Open ASR Leaderboard's seven English test sets—competitive with OpenAI's Whisper large-v3, which requires a 9.26 GB download and delivers 6.04 percent error. For comparison, NVIDIA's Parakeet itself is 2.51 GB. Bark 3.0 Pro, a larger variant in the family, ranks first on the Open ASR Leaderboard with 4.286 percent error, outperforming other open-source models.