ARK-ASR-3B

ARK-ASR-3B

Turns speech into text in many languages, with a Whisper-style audio encoder feeding a Qwen decoder. For developers and researchers who want multilingual transcription, with English and Chinese results reported.

Made by Edge0 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: A speech recognition model that turns audio into text in the many languages the card lists. It joins a Whisper-style audio encoder, a small adapter and a Qwen decoder, and it relies on custom code shipped with the model. The card reports an average word error rate of 5.04% on the Hugging Face Open ASR Leaderboard's English short-form set.

Who it suits: Developers and researchers who need accurate multilingual transcription and are comfortable running Python with transformers.

Industry | Who uses it | How they use it | Example
Media | Podcast and video teams | Transcribe multilingual recordings into text | Captions from a Chinese or English interview
Finance | Analysts | Turn recorded calls into searchable text | A transcript of a recorded briefing
Software | Voice feature developers | Add speech input to an app through transformers or vLLM | A transcription service behind an app
Research | Speech researchers | Compare error rates against leaderboard results | Check a new dataset against the card's numbers

You can skip it if you need a language outside the card's list or a small download, since the BF16 weights take about 8.1 GB.

Where to use it: It runs in Python with the transformers library on a GPU or a CPU, and the card also points to a vLLM serving script.

• Python with transformers and torch, on Windows, macOS or Linux
• A CUDA GPU when you have one, since the card's code falls back to the CPU otherwise
• vLLM, using the serving script in the project the card names

Use case | Who | What they get back
Multilingual transcription | Media teams | Text from audio in any language the card lists
Speech service | Developers | A transcription endpoint through the vLLM script
Benchmark checks | Researchers | Error rates to compare against other models

How to set it up: Fetch the files, install the Python packages, and load the model with remote code allowed. The card's own inference script is the recommended way to run it.

• Python with transformers and torch
• Audio at a 16 kHz sampling rate
• Room for about 8.1 GB of weights in BF16

1. Install the Hugging Face tools and download the model files.
2. Get the card's inference script from the AutoArk open-audio-opd project.
3. Load the model with trust_remote_code set to True.
4. Run it on a 16 kHz audio file and read the transcript.

pip install -U huggingface_hub
hf download Edge0/ARK-ASR-3B

How to use it: Load the processor, tokenizer and model, then feed in your audio. The card's loading code is below, and it uses the path AutoArk-AI/ARK-ASR-3B and a sample file named libai.wav.

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer

model_path = "AutoArk-AI/ARK-ASR-3B"
audio_path = "assets/libai.wav"

device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch_dtype,
    attn_implementation="sdpa",
).to(device)
model.eval()

Tool | What it does
ark_asr_transformers.py | The card's recommended inference script, which prepares the audio prompt and cleans up the output
ark_asr_vllm | Scripts for serving the model under vLLM
trust_remote_code=True | The option that lets transformers load the model's custom arkasr code
AutoProcessor | Loads the audio processor that prepares your recording

Example: Point the card's script at a 16 kHz recording, such as the sample file libai.wav it names. The script loads the model, runs the audio through it and prints the written transcript.

For newcomers and experts: Speech recognition turns spoken audio into written text. A first small try is to run the card's script on a short clip and read what comes out, before you test your own recordings.

For experts, a Whisper-style encoder with RoPE turns audio into features, an MLP adapter merges them, and they replace audio placeholder embeddings inside a Qwen decoder before the transcript is generated. The custom arkasr code means you run code from the repository, so read it first. The 3B model beats the card's 0.6B sibling on every benchmark column shown, at the cost of a larger download.

Record or load 16 kHz audio
Whisper-style encoder makes features
MLP adapter merges them
Qwen decoder reads placeholder embeddings
Transcript text comes out

Good to know: The card reports these English word error rates for ARK-ASR-3B, where lower is better.

Title: English word error rate, lower is better
AMI: 8.79%
Earnings22: 8.23%
GigaSpeech: 6.98%
LS Clean: 1.03%
LS Other: 2.35%
SPGISpeech: 2.46%
VoxPopuli: 5.47%
Average: 5.04%

• Languages: the card lists many, from Chinese and English to Lithuanian.
• Remote code: loading needs trust_remote_code set to True, so review the code before you run it.
• Chinese results: the card reports error rates of 1.80% on AISHELL-1, 4.97% on WenetSpeech test meeting and 4.58% on WenetSpeech test-net.
• Speed: the card reports an RTFx of 490.98 on the leaderboard set.
• License: Apache-2.0.
• Cost to run: none named, no API key, account or paid service.
• Data it reaches: it reads the audio files you give it and downloads the model files from Hugging Face.
• Not for: the card names no limits, and its state-of-the-art claim covers only the English short-form set.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 4.1B
• Weights: about 8.1 GB in BF16; about 2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,655
• Likes: 115
• License: apache-2.0
• Languages: zh, en, de, ja, fr, ko, es, pl
• Quantized versions: 5
• Fine-tunes: 4
• Tags: transformers, safetensors, arkasr, text-generation, automatic-speech-recognition, speech, audio, pytorch, vllm, ark-asr

Links

• Model: https://huggingface.co/Edge0/ARK-ASR-3B
• Files: https://huggingface.co/Edge0/ARK-ASR-3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Edge0/ARK-ASR-3B

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