---
license: cc-by-4.0
language:
  - en
library_name: pytorch
pipeline_tag: automatic-speech-recognition
tags:
  - automatic-speech-recognition
  - parakeet
  - tdt
  - safetensors
base_model:
  - ConwayResearch/Underdog-Bark-3.0-Plus
  - ConwayResearch/Underdog-Bark-3.0-Clean
  - ConwayResearch/Underdog-Bark-3.0-Pro
---

# Underdog Bark 3.0 — MAESTRO mirror

Mirror of the three [Underdog Bark 3.0](https://huggingface.co/ConwayResearch) English speech-recognition
checkpoints that MAESTRO ships, with the NVIDIA feature-extractor and tokenizer files they share, so the
in-app downloader has one stable, self-contained, offline-capable source.

| folder | source | what it is |
|---|---|---|
| `plus/` | [ConwayResearch/Underdog-Bark-3.0-Plus](https://huggingface.co/ConwayResearch/Underdog-Bark-3.0-Plus) | verbatim — every word as spoken, fillers kept |
| `clean/` | [ConwayResearch/Underdog-Bark-3.0-Clean](https://huggingface.co/ConwayResearch/Underdog-Bark-3.0-Clean) | fillers, false starts and repeats removed |
| `pro/` | [ConwayResearch/Underdog-Bark-3.0-Pro](https://huggingface.co/ConwayResearch/Underdog-Bark-3.0-Pro) | the larger-download verbatim checkpoint |
| `processor/` | [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) | feature extractor + tokenizer (shared) |

Each variant folder is a standard `ParakeetForTDT` checkpoint (`config.json`, `generation_config.json`,
`model.safetensors`). Exact source revisions and the sha256 of every file are in `sha256.json`.

## Attribution (CC-BY-4.0)

Underdog Bark 3.0 is © **Conway Research**, CC BY 4.0. It derives from **nvidia/parakeet-tdt-0.6b-v3**
(© NVIDIA Corporation, CC BY 4.0) by way of **moondream/parakeet-ultra** (© Moondream, CC BY 4.0).
License: https://creativecommons.org/licenses/by/4.0/ — full chain in `NOTICE`.

## Changes in this mirror

- `plus/` and `clean/`: Conway publishes these as an entropy-coded `model.bark` that needs the
  `constriction` ANS decoder and transformers ≥ 5.19 to load. They were decoded once with Conway's own
  decode routine and written as safetensors. The decode is exact and so is the storage: each tensor is
  fp16 where that is bit-identical to the decoded value (the level-coded encoder weights, 604M params)
  and fp32 otherwise (23M params). No values were changed.
- `pro/` and `processor/`: byte-identical copies.

## Use

MAESTRO loads these with a vendored Parakeet-TDT graph (AI › Transcribe › Underdog Bark). With
transformers ≥ 5.19 the folders also load directly:

```python
from transformers import AutoProcessor, ParakeetForTDT
model = ParakeetForTDT.from_pretrained("AEmotionStudio/underdog-bark-models", subfolder="plus")
processor = AutoProcessor.from_pretrained("AEmotionStudio/underdog-bark-models", subfolder="processor")
```

Prepend 0.25 s of silence for Plus and Pro (Conway's evaluated setting; Clean uses none). Prefer fp16 or
fp32: in MAESTRO's parity runs bf16 changed the transcript on 3 of 10 clips for Pro.