README.md

---
license: apache-2.0
base_model: Audio8/ARK-ASR-0.6B
base_model_relation: quantized
library_name: onnx-asr
pipeline_tag: automatic-speech-recognition
tags:
  - onnx
  - onnxruntime
  - automatic-speech-recognition
  - speech
  - audio
  - asr
  - int8
  - speech-llm
language:
  - zh
  - en
  - de
  - ja
  - fr
  - ko
  - es
  - pl
  - it
  - ro
  - hu
  - cs
  - nl
  - fi
  - hr
  - sk
  - sl
  - et
  - lt
---

# ARK-ASR-0.6B ONNX

ONNX export of [Audio8/ARK-ASR-0.6B](https://huggingface.co/Audio8/ARK-ASR-0.6B) for
[onnx-asr](https://github.com/istupakov/onnx-asr). All credit for the model goes to
Audio8 (AutoArk AI). This repository only contains the converted graphs; the
weights are the original ones.

The model is a speech-LLM: a Whisper-large-v3-style audio encoder with rotary
position embeddings, an MLP adapter that merges four encoder frames into one
embedding, and a Qwen2 0.6B causal language model that writes the transcription.

## Usage

```sh
pip install onnx-asr[cpu,hub]
```

```py
import onnx_asr

model = onnx_asr.load_model("speech-llm", "OpenVoiceOS/ARK-ASR-0.6B-onnx")
print(model.recognize("audio.wav"))

# int8 weights
model = onnx_asr.load_model("speech-llm", "OpenVoiceOS/ARK-ASR-0.6B-onnx", quantization="int8")
```

## Files

| File | Contents |
| --- | --- |
| `encoder.onnx` | audio encoder and MLP adapter, log-mel features in, LM embeddings out |
| `embed_tokens.onnx` | token embedding table |
| `decoder.onnx` | Qwen2 decoder with KV cache, logits out |
| `*.int8.onnx` | dynamically quantized int8 weights |
| `config.json` | model type, prompt token ids, suppressed token ids |
| `vocab.json` | tokenizer vocabulary for detokenization |

## Graph contract

| Graph | Inputs | Outputs |
| --- | --- | --- |
| `encoder.onnx` | `input_features (1, 128, frames)` | `audio_embeds (1, frames/8, 896)` |
| `embed_tokens.onnx` | `input_ids (1, S)` | `inputs_embeds (1, S, 896)` |
| `decoder.onnx` | `inputs_embeds (1, S, 896)`, `attn_bias (1, 1, S, P+S)`, `position_ids (1, S)`, `past_key_values.{0..23}.{key,value} (1, 2, P, 64)` | `logits (1, S, 163958)`, `present.{0..23}.{key,value} (1, 2, P+S, 64)` |

## Accuracy

Four FLEURS clips (2 English, 2 Mandarin), greedy decoding, compared against the
PyTorch model in float32:

* fp32 ONNX: 4 of 4 transcriptions identical to PyTorch, character for character.
* int8 ONNX: 3 of 4 identical. One Mandarin clip stops early and loses the last
  clause.

Speed on a 12-core CPU under load: RTF 0.35 to 0.61 (fp32) and 0.09 to 0.29
(int8).

## Licence

Apache 2.0, the same licence as the source model. The model was published by
Audio8; see the [source repository](https://huggingface.co/Audio8/ARK-ASR-0.6B)
and the paper [arXiv:2605.28139](https://arxiv.org/abs/2605.28139).