strixAE

strixAE is an audio-language reasoning agent obtained by applying Group Relative Policy Optimization (GRPO) to Audio-Reasoner and merging the trained weights into a standalone checkpoint. It uses the Qwen2AudioForConditionalGeneration architecture and is intended for audio understanding, structured reasoning, and audio-restoration pipeline decisions.

Status: This repository contains a research checkpoint. Independent benchmark results for this exact GRPO checkpoint have not yet been published.

Model details

Item	Value
Model type	Audio-language conditional generation
Architecture	Qwen2AudioForConditionalGeneration
Parameters	8.40B
Weight format	BF16 Safetensors (4 shards)
Context length	8,192 text positions
Base checkpoint	zhifeixie/Audio-Reasoner
Post-training	GRPO reinforcement learning

Intended use

The model is designed for research involving:

• audio understanding across speech, music, and environmental sounds;
• audio-quality and degradation analysis;
• deciding whether candidate restoration operations should be applied or skipped;
• selecting and ordering an audio-restoration pipeline;
• structured audio reasoning and question answering.

For restoration tasks, prompts should explicitly provide the available operations and request a decision for every operation. A useful output contract is:

<THINK>
Analyze the audio and evaluate every candidate operation.
</THINK>

**Detected Audio Issues:**
- ...

**Restoration Sequences Analysis:**
1. denoise(...)
   - Decision: APPLY / SKIP
   - Reason: ...

**Selected Restoration Pipeline:**
1. ...

Usage with Transformers

Install the runtime dependencies:

pip install "transformers>=4.57.0" accelerate librosa soundfile

Run inference on a local audio file:

import librosa
import torch
from transformers import AutoProcessor, Qwen2AudioForConditionalGeneration

model_id = "wujunjiehhs/strixAE"
audio_path = "example.wav"

processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2AudioForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

conversation = [
    {
        "role": "system",
        "content": (
            "You are an audio reasoning assistant. Analyze the audio carefully "
            "and provide a clear, structured answer."
        ),
    },
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio": audio_path},
            {
                "type": "text",
                "text": (
                    "Identify the audio issues. Evaluate denoising, dereverberation, "
                    "source separation, and super-resolution. For each operation, "
                    "return APPLY or SKIP with a reason, then propose an ordered pipeline."
                ),
            },
        ],
    },
]

prompt = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=False,
)
audio, _ = librosa.load(
    audio_path,
    sr=processor.feature_extractor.sampling_rate,
    mono=True,
)
inputs = processor(
    text=prompt,
    audio=audio,
    sampling_rate=processor.feature_extractor.sampling_rate,
    return_tensors="pt",
    padding=True,
)
inputs = {k: v.to(model.device) for k, v in inputs.items()}

with torch.inference_mode():
    generated_ids = model.generate(**inputs, max_new_tokens=1024)

generated_ids = generated_ids[:, inputs["input_ids"].shape[1]:]
response = processor.batch_decode(
    generated_ids,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0]
print(response)

The repository includes its chat template, tokenizer, processor configuration, generation configuration, and merged model weights; no adapter merge is required at inference time.

Training information

This checkpoint was produced by GRPO post-training and then merged for deployment. Exact training data composition, reward functions, hyperparameters, compute, and checkpoint-selection criteria are not included in this release. Users should not assume that benchmark numbers reported for the upstream Audio-Reasoner checkpoint transfer unchanged to this model.

Limitations and responsible use

• The model may hallucinate acoustic events or infer degradations that are not present.
• Restoration recommendations are model-generated judgments, not objective signal measurements.
• Structured output is not guaranteed; validate the response before using it in an automated pipeline.
• Performance may degrade for very long audio, low-resource languages, unseen codecs, unusual sampling rates, or out-of-domain recordings.
• Do not use the model as the sole basis for safety-critical, medical, legal, forensic, surveillance, or high-impact decisions.
• Audio can contain personal or sensitive information. Ensure that collection, processing, and sharing comply with applicable consent, privacy, copyright, and data-protection requirements.

中文简介

strixAE 是在 Audio-Reasoner 基础上进行 GRPO 强化学习得到的音频理解代理。它面向音频理解、音质问题分析、候选修复操作的 APPLY/SKIP 判断，以及修复流程排序。当前仓库尚未发布该 GRPO 检查点的独立评测结果，建议在实际数据上验证后再部署。

Acknowledgements and citation

This work builds on Audio-Reasoner and Qwen2-Audio. Please cite the upstream Audio-Reasoner work when using this checkpoint:

@misc{xie2025audioreasonerimprovingreasoningcapability,
  title        = {Audio-Reasoner: Improving Reasoning Capability in Large Audio Language Models},
  author       = {Zhifei Xie and Mingbao Lin and Zihang Liu and Pengcheng Wu and Shuicheng Yan and Chunyan Miao},
  year         = {2025},
  eprint       = {2503.02318},
  archivePrefix= {arXiv},
  primaryClass = {cs.SD},
  url          = {https://arxiv.org/abs/2503.02318}
}

License

This release follows the MIT license included with the upstream Audio-Reasoner project. Users are responsible for reviewing and complying with the terms of all upstream models, datasets, and dependencies.