Compact English speech-to-text model family from Underdog that runs on-device on Mac, starting at 163 MB.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Bark 3.0 is the speech-recognition model family used in Underdog, a Mac app for dictation (hold fn and talk) and meeting transcription. Transcription runs locally on the Mac and the vendor states nothing is sent anywhere. The smallest model is a 163 MB download.
The family has four variants: Bark 3.0 (every word, 163 MB), Bark 3.0 Plus (every word, fewer errors), Bark 3.0 Clean (about 180 MB, removes fillers, false starts and repeats, and is the default in Underdog) and Bark 3.0 Pro (full precision, 1.25 GB, Hugging Face only). On the Open ASR Leaderboard's English sets, the vendor reports average word error rates of 4.94% (Bark 3.0), 4.76% (Plus) and 4.29% (Pro), and says it is comparable to NVIDIA Parakeet TDT 0.6B v3 and ahead of Whisper large-v3. These figures are self-reported.
Limits: English only (Underdog offers Whisper Large v3 Turbo for other languages), no word timings or speaker names yet, and the custom words list does not work with Bark 3.0 yet. The vendor says it is built on NVIDIA parakeet-tdt-0.6b-v3 and Moondream's parakeet-ultra, both CC-BY 4.0. A YouTube reviewer said the license allows commercial use with attribution.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Tiny speech-to-text transcription model family (163 MB smallest, plus and pro versions up to 1.25 GB). Creator tested it in a Hugging Face space. Competes with Whisper Large v3 and Parakeet; license allows commercial use with attribution.
Closest neighbours by meaning, not just by category.
Model · ServiceNow
Open-weights 27B speech-to-text model from ServiceNow that adds a speech interface to a Gemma 4 backbone, with tool calling.
Model · Cactus
Compact 16.9 MB speech-to-text model from Cactus that runs on CPU with no dependencies.
Model · Fermion Research
Open 164 MB English speech recognition model from Fermion Research, derived from Parakeet TDT 0.6B v3.
Model · Nyra Labs
Controllable multilingual speech recognition with verbatim or clean output and word-level timestamps, from Nyra Labs.
Model · Edge0
Open-weights 0.17B-parameter multilingual text-to-speech model with zero-shot voice cloning, runs on CPU via ONNX.
Model · Hao AI Lab
Open-weight 4-step sparse-distilled version of MiniMax H3 for fast text-to-video-and-audio generation on Blackwell GPUs.
