Underdog Bark 3.0

Underdog Bark 3.0

Turns English speech into text with a 0.6B model in a 163 MB download, built on NVIDIA's Parakeet TDT 0.6B v3. For developers and Apple-silicon users who want small local transcription. Non-commercial license.

Made by ConwayResearch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Bark 3.0 is an English speech recognition model with 0.6B parameters, built on NVIDIA's Parakeet TDT 0.6B v3. The download is only 163 MB. Its authors report an average word error of 4.94 percent across seven English test sets from the Open ASR Leaderboard.

Who it suits: Developers and Mac users who want accurate English transcription from a small file that can run on their own device. It writes what was said, fillers included, so it suits verbatim transcripts.

Industry | Who uses it | How they use it | Example
Meetings | Teams | Turn recorded meetings into text | Run a recording through the model and read the transcript
App development | App developers | Run speech recognition on the device | Use the model inside Underdog on Apple silicon
Speech research | ML researchers | Compare small recognition models | Score it against Phonon-2 on the same test sets
Media | Editors | Write out spoken audio | Transcribe a recorded talk in English

You can skip it if you need languages other than English, commercial use, or dictation that drops the um and uh sounds.

Where to use it:

• Underdog, which runs the model with no Python, using a native Swift runtime on MLX
• Python on Windows, macOS or Linux, through the transformers library the listing names
• mlx-audio on an Apple computer, which the card used for its speed test

Use case | Who | What they get back
Transcription | Developers | English text from speech, fillers included
On-device dictation | Mac users | Text from your voice without sending audio away
Benchmarking | Researchers | Word error rates to compare against other models

How to set it up: You need Python and the Hugging Face tool. Underdog needs an Apple-silicon Mac.

• Python on Windows, macOS or Linux
• About 163 MB of disk space for the download
• An Apple-silicon Mac, if you want to run it in Underdog

1. Install Python if you do not have it.
2. Install the Hugging Face hub tool, in the version the card pins.
3. Download the model files.
4. Follow the card's Run it section for the rest, which continues past what is shown here.

pip install huggingface_hub==2.2.0
hf download ConwayResearch/Underdog-Bark-3.0

How to use it: In Underdog you pick the model and it reads the file directly. In Python, follow the card's Run it section after the download. For dictation that leaves out the fillers, the card points to a sibling called Underdog Bark 3.0 Clean.

Tool | What it does
Underdog | Runs the model file with a native Swift runtime on MLX, with no Python
mlx-audio | A tool the card used to time transcription on an Apple M2 laptop
Underdog Bark 3.0 Clean | A sibling model for dictation that leaves out the fillers

Example: You record a meeting and run it through the model. The text comes back as spoken, with the um and uh sounds left in. On an Apple M2 laptop the card measured about 1.5 seconds for one minute of audio.

For newcomers and experts: Speech recognition turns spoken audio into written text. Word error rate is the share of words the model gets wrong, so lower is better. A first small step is to download the 163 MB files and read the card's Run it section.

The card compares it with larger models on seven English sets. It reports 4.937 against 4.941 for Parakeet TDT 0.6B v3 at three decimals, and 6.04 for Whisper large-v3. The Bark rows use the leaderboard's scoring code but are not on the leaderboard yet. Underdog reads the model file into memory in roughly 2 seconds on Apple silicon. It then checks each download against a fixed size and a SHA-256 hash.

Download the 163 MB model
Load it in Underdog or in Python
Feed it English audio
Get a transcript with fillers kept

Title: Average word error in percent on seven English sets, lower is better
Bark 3.0: 4.94%
Parakeet TDT 0.6B v3: 4.94%
Qwen3-ASR-0.6B: 5.16%
Phonon-2: 5.21%
Whisper large-v3: 6.04%

Good to know:

• It is an English model and it does not clean up speech, so fillers stay in the text.
• The accuracy numbers are the authors' own, scored with the leaderboard's code but not yet listed there.
• In Underdog the card says your audio never leaves the device.
• Download size is 163 MB, against 2.5 GB for the Parakeet model it builds on.
• License: cc-by-nc-4.0, which allows non-commercial use only.
• Cost to run: none named, you use your own hardware.
• Data it reaches: your audio, plus files downloaded from Hugging Face. Underdog keeps audio on the device.
• Not for: commercial use, or speech in languages other than English.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Downloads in the last 30 days: 413
• Likes: 95
• License: cc-by-nc-4.0
• Languages: en
• Based on: nvidia/parakeet-tdt-0.6b-v3
• Quantized versions: 5
• Fine-tunes: 3
• Tags: transformers, coreml, underdog, speech-to-text, low-bit, on-device, apple-silicon, parakeet, automatic-speech-recognition

Links

• Model: https://huggingface.co/ConwayResearch/Underdog-Bark-3.0
• Files: https://huggingface.co/ConwayResearch/Underdog-Bark-3.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ConwayResearch/Underdog-Bark-3.0

Then follow the usage in its model card. Nothing in this kit runs by itself.