ModelPrésentéExterne

Azeros

AZeroS (Auden Zero-instruction-tuned Speech-LLM) extends a frozen LLM to speech via Self-Generated Instruction-Free Tuning (SIFT). It keeps the LLM and audio encoders frozen and trains lightweight projection modules...

Sourcehuggingface.co/AudenAI/azeros/tree/e9392d32639e7079cb0e24f7373207ee658a8baf

TypeOtherModèle de baseQwen/Qwen2.5-7B-InstructVersionRevision e9392d3BClasse

Source du modèle

Type: Other
Taille du fichier: 182 MB–4,59 GB
Modèle de base: Qwen/Qwen2.5-7B-Instruct
Version: vRevision e9392d3
Mis à jour: 25 oct. 2026
Téléchargements: 21
Source: Hugging Face

Description de la source

AZeroS

AZeroS (Auden Zero-instruction-tuned Speech-LLM) extends a frozen LLM to speech via Self-Generated Instruction-Free Tuning (SIFT). It keeps the LLM and audio encoders frozen and trains lightweight projection modules on speech–text pairs, achieving strong semantic and paralinguistic performance with modest training cost, generalizing well to unseen instructions.

🔗 Paper: https://arxiv.org/pdf/2601.06086
🔗 Code: https://github.com/AudenAI/Auden/tree/main/examples/azeros
🔗 Model: https://huggingface.co/AudenAI/azeros
🔗 Auden Repo: https://github.com/AudenAI/Auden

🔍 What Can This Model Do?

🎙️ Speech understanding (semantic content understanding and dialog)
😊 Paralinguistic analysis (emotion, age, gender, etc.)

Quick Start

import torch
from model import AZerosModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AZerosModel.from_pretrained("AudenAI/azeros").to(device)

wav_files = ["speech1.wav", "speech2.wav"]
messages = [
    [
        {
            "role": "user",
            "content": f"{model.audio_token_wrapped} Please analyze speech content and paralinguistic information.",
        }
    ]
    for _ in wav_files
]

generate_config = {
    "max_new_tokens": 200,
    "num_beams": 1,
    "do_sample": False,
    "min_length": 1,
    "repetition_penalty": 1.0,
    "length_penalty": 1.0,
    "top_p": None,
    "top_k": None,
    "temperature": None,
}

outputs = model.generate(wav_files, messages, **generate_config)
print(outputs)

Auden Setup (Required)

This model relies on the Auden codebase for loading and inference:

git clone https://github.com/AudenAI/Auden.git
cd Auden
pip install -e .
cd examples/azeros

📌 Model Characteristics

Input: Raw audio waveform (16 kHz) or text
Output: Text responses regarding to the input
Backend LLM: Qwen2.5-7B-Instruct
Encoders: TTA and Auden-Voice
Architecture: Frozen LLM + frozen audio encoders + lightweight projection modules
Training paradigm: Self-Generated Instruction-Free Tuning (SIFT)

📊 Evaluation

VoiceBench

Model	Alpaca Eval	Comm Eval	Wild Voice	SD-QA	BBH	Adv Bench	IF Eval	OBQA	MMSU	Overall
Text Only Model
Qwen2.5	4.66	4.55	4.62	62.03	80.00	99.04	70.14	84.84	71.57	82.69
Qwen2.5 (TN)	4.61	4.53	4.56	63.84	56.30	98.85	66.11	74.07	64.51	77.52
Cascaded System
Whisper+GPT-4o	4.80	4.47	4.62	75.77	87.20	98.27	76.51	92.97	81.69	87.80
Whisper+Qwen2.5	4.64	4.33	4.21	58.50	52.85	98.27	63.99	78.24	69.00	76.05
End-to-end Speech-LLM
GPT-4o	4.78	4.49	4.58	75.50	84.10	98.65	76.02	89.23	80.25	86.75
Moshi	2.01	1.60	1.30	15.64	47.40	44.23	10.12	25.93	24.04	29.51
Phi-4-multimodal	3.81	3.82	3.56	39.78	61.80	100.00	45.35	65.93	42.19	64.32
GLM-4-Voice	3.97	3.42	3.18	36.98	52.80	88.08	25.92	53.41	39.75	56.48
Qwen2-Audio	3.42	3.29	2.76	31.65	53.00	99.04	26.35	48.35	36.14	53.77
DeSTA2.5	3.73	2.52	3.30	46.47	62.40	97.69	65.47	72.75	58.56	66.04
Qwen2.5-Omni	3.88	3.77	3.52	46.75	63.70	97.31	40.19	81.54	61.45	68.26
Qwen3-Omni-30B	4.74	4.54	4.58	76.90	80.40	99.30	77.80	89.70	68.10	85.49
AZeroS (ours)	4.44	4.18	3.91	60.22	56.30	98.65	61.29	72.09	59.01	73.13

AIRBench

Model	Gender	Emotion	Age	LID	Entity	Intent	Avg	Chat
Cascaded System
Whisper+GPT-4o	21.90	59.50	41.10	96.80	69.80	87.70	62.80	7.54
Whisper+Qwen2.5	28.36	50.80	36.40	88.00	73.60	82.70	59.98	7.34
End-to-end Speech-LLM
GPT-4o	*	49.10	*	76.00	61.60	85.80	*	7.53
Gemini2.5-pro	90.70	60.70	34.10	99.10	68.50	92.20	74.22	8.52
SALMONN	35.50	29.90	48.70	28.10	51.70	36.70	38.43	6.16
GLM-4-Voice	23.91	22.95	18.70	25.40	27.90	21.10	23.33	5.53
Qwen2-Audio	64.71	48.15	23.10	77.80	87.00	84.70	64.24	7.20
DeSTA2.5	84.24	64.30	65.60	97.30	65.20	83.70	76.72	7.57
Qwen2.5-Omni	89.76	54.85	44.80	89.70	79.70	88.60	74.57	6.97
Qwen3-Omni-30B	91.11	62.20	36.90	97.70	80.40	90.70	76.50	7.85
AZeroS (ours)	86.75	71.45	61.30	84.80	73.60	85.60	77.25	8.28

An additional prompt is added to ensure steady output of choices: “Please make your choice among A/B/C/D and do not output other texts.”

⚠️ Limitations

Trained on public datasets; performance may degrade on out-of-domain audio.
Not designed for safety-critical applications.

Citation

If you use AZeroS in your research, please cite:

@article{shao2026azeros,
  title={AZEROS: Extending LLM to Speech with Self-Generated Instruction-Free Tuning},
  author={Shao, Yiwen and Liu, Wei and Li, Jiahong and Wang, Tianzi and Wei, Kun and Yu, Meng and Yu, Dong},
  journal={arXiv preprint arXiv:2601.06086},
  year={2026}
}

Sources

1 source

Hugging FacePrincipalVersion e9392d32639e7079cb0e24f7373207ee658a8baf

huggingface.co/AudenAI/azeros/tree/e9392d32639e7079cb0e24f7373207ee658a8baf

Vérifié 25 oct.

Download files

Artefacts du modèle

7 artefacts

llm/model-00001-of-00004.safetensors

safetensors · 4,54 GB · SHA-256 69f259295cfa…f6c9 · Hugging Face

Download shard

llm/model-00002-of-00004.safetensors

safetensors · 4,59 GB · SHA-256 960c02015eb7…37e6 · Hugging Face

Download shard

llm/model-00003-of-00004.safetensors

safetensors · 4,03 GB · SHA-256 ced2cfa9d180…ec2f · Hugging Face

Download shard

llm/model-00004-of-00004.safetensors

safetensors · 1,02 GB · SHA-256 053d06aa9a65…02e8 · Hugging Face

Download shard

paraling_encoder/model.safetensors

safetensors · 583 MB · SHA-256 00a27891e64a…6e17 · Hugging Face

Download file

pretrained.pt

pt · 182 MB · SHA-256 15fda43fd3f0…457c · Hugging Face

Download file

speech_encoder/model.safetensors

safetensors · 583 MB · SHA-256 cbf2927b5552…eadc · Hugging Face

Download file

Extraits de sources

3 extraits

Readme7 aoûthuggingface.co/AudenAI/azeros

--- license: apache-2.0 datasets: - wenetspeech - gigaspeech - common_voice - iemocap - crema-d - meld - ravdess - tess - dailytalk - aishell-1 - emotiontalk - cs-dialogue - voxceleb2 language: - en - zh base_model: - Qwen/Qwen2.5-7B-Instruct pipeline_tag: audio-text-to-text tags: - speech - speech-llm - audio - instruction-free - paralinguistic --- # AZeroS **AZeroS** (Auden Zero-instruction-tuned Speech-LLM) extends a frozen LLM to speech via **Self-Generated Instruction-Free Tuning (SIFT)**. It keeps the LLM and audio encoders frozen and trains lightweight projection modules on speech–text pairs, achieving strong semantic and paralinguistic performance with modest training cost, generalizing well to unseen instructions. 🔗 **Paper**: https://arxiv.org/pdf/2601.06086 🔗 **Code**: https://github.com/AudenAI/Auden/tree/main/examples/azeros 🔗 **Model**: https://huggingface.co/AudenAI/azeros 🔗 **Auden Repo**: https://github.com/AudenAI/Auden ## 🔍 What Can This Model Do? - 🎙️ **Speech understanding** (semantic content understanding and dialog) - 😊 **Paralinguistic analysis** (emotion, age, gender, etc.) ## Quick Start ```python import torch from model import AZerosModel device = "cuda" if torch.cuda.is_available() else "cpu" model = AZerosModel.from_pretrained("AudenAI/azeros").to(device) wav_files = ["speech1.wav", "speech2.wav"] messages = [ [ { "role": "user", "content": f"{model.audio_token_wrapped} Please analyze speech content and paralinguistic information.", } ] for _ in wav_files ] generate_config = { "max_new_tokens": 200, "num_beams": 1, "do_sample": False, "min_length": 1, "repetition_penalty": 1.0, "length_penalty": 1.0, "top_p": None, "top_k": None, "temperature": None, } outputs = model.generate(wav_files, messages, **generate_config) print(outputs) ``` ## Auden Setup (Required) This model relies on the Auden codebase for loading and inference: ```bash git clone https://github.com/AudenAI/Auden.git cd Auden pip install -e . cd examples/azeros ``` ## 📌 Model Characteristics - Input: Raw audio waveform (16 kHz) or text - Output: Text responses regarding to the input - Backend LLM: Qwen2.5-7B-Instruct - Encoders: [TTA](https://huggingface.co/AudenAI/auden-encoder-tta-m10) and [Auden-Voice](https://huggingface.co/AudenAI/auden-encoder-voice) - Architecture: Frozen LLM + frozen audio encode...

Note25 oct.huggingface.co/AudenAI/azeros

Source context: 21 downloads · 2 likes · Pipeline audio-text-to-text · Repo AudenAI/azeros

Note7 aoûthuggingface.co/AudenAI/azeros

Source context: 6 downloads · 2 likes · Pipeline audio-text-to-text · Repo AudenAI/azeros