Audio8-ASR

Audio8-ASR-0.1B is a compact autoregressive ASR model whose language-model component has only 0.1B parameters. It supports multilingual speech recognition for languages including Chinese, English, French, German, Japanese, Korean, and Cantonese. We position it as one of the smallest usable performance ASR models in the LLM era.

This base repository provides the Hugging Face Transformers checkpoint. We also provide deployment-focused releases:

• Audio8-ASR-0.1B-onnx-runtime
• Audio8-ASR-0.1B-iOS-ANE

The ONNX Runtime release is designed for edge-device deployment and can run with roughly 1.1 GB peak memory footprint, depending on device, runtime configuration, and workload.

The iOS release is designed for local iPhone transcription with roughly 200 MB peak runtime memory footprint, depending on device, iOS version, and workload.

Demo Video

Your browser does not support the video tag.

Evaluation Results

The Open ASR results use the seven current public splits from hf-audio/open-asr-leaderboard at dataset revision b6bdcd0beb34f8975dc659796176d88f43aff502. They were measured with the standalone Transformers package on standardized H200 Hugging Face Jobs using BF16, eager attention, greedy decoding, max_new_tokens=256, and the documented 30-second audio cap. Per-split batch sizes were 1152, 1024, 1408, 1024, 1024, 2048, and 628. Raw manifests are stored in hf://buckets/AutoArk-AI/audio8-asr-open-asr-results, and the corresponding machine-readable results are provided in .eval_results/open_asr_leaderboard.yaml.

The internal canonical WenetSpeech results come from the reproducibility-checked teacher0p6B-step3000 export with batch size 128. Its effective model tensors are byte-identical to this standalone release; the release only removes a redundant tied LM-head tensor and packages the same weights for standalone use. Chinese results are reported as character error rate. AISHELL is intentionally excluded from this table.

Model Overview

• Task: automatic speech recognition
• Checkpoint format: safetensors
• Sampling rate: 16 kHz
• Decoder: 8-layer Qwen-style causal LM
• Audio front end: Qwen3-ASR audio encoder plus MLP adapter/projector
• Language-model parameters: 103,502,336 (about 0.104B)
• End-to-end unique parameters: 323,990,528 (about 0.324B)
• Runtime: Hugging Face Transformers
• Hotwords: optional decode-time logit boosting, no fine-tuning required

The model should be loaded with trust_remote_code=True.

Files

• config.json, tokenizer files, processor files, and model.safetensors
• configuration_arkasr.py, modeling_arkasr.py, processing_arkasr.py
• qwen3_asr_audio_config.py, qwen3_asr_audio_model.py
• hotword/: backend-agnostic hotword trie
• examples/: Transformers inference examples

The root config.json is intentionally kept in this repository so Hugging Face can recognize the model package and count downloads through normal model-file queries.

Transformers Inference

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_path = "AutoArk-AI/Audio8-ASR-0.1B"
audio_path = "path/to/audio.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)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch_dtype,
    attn_implementation="eager",
).to(device)
model.eval()

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "audio", "path": audio_path},
            {"type": "text", "text": "Please transcribe this audio."},
        ],
    }
]

batch = processor.apply_chat_template(
    conversation,
    return_tensors="pt",
    sampling_rate=16000,
    audio_padding="longest",
    add_generation_prompt=True,
    audio_max_length=30 * 16000,
    text_kwargs={"padding": "longest", "truncation": True, "max_length": 1000},
)
batch = {key: value.to(device) if hasattr(value, "to") else value for key, value in dict(batch).items()}

with torch.inference_mode():
    output_ids = model.generate(**batch, max_new_tokens=128, do_sample=False)

prompt_len = int(batch["input_ids"].shape[1])
text = processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True).strip()
print(text)

Equivalent script:

python examples/transcribe.py path/to/audio.wav --model AutoArk-AI/Audio8-ASR-0.1B

For local staging before upload:

python examples/transcribe.py path/to/audio.wav --model .

Hotword Boosting

Hotwords are applied at decode time by nudging logits for tokenizer paths that match the requested words. Th