Aero-1-Audio is a lightweight audio model developed by LMMs-Lab, built on Qwen-2.5-1.5B, with only 150 million parameters. Designed for long audio processing, supporting 15 minutes of continuous audio input without segmentation. It has high accuracy in speech recognition (ASR) tasks and has the lowest word error rate on AMI, LibriSpeech, SPGISpeech and other data sets.

Aero-1-Audio

Core parameters and statistics

Parameters	Official verifiable information
Product positioning	Lightweight audio model for long-form audio
Basic model	Qwen-2.5-1.5B
Parameter amount	150 million (150M)
Maximum audio length	15 minutes of continuous input
Training data	~5 billion tokens (50,000 hours of audio)
Training resources	16 H100 GPUs, completed in 1 day
FLOP utilization	0.34 (optimized with sequence packing)
Publishing platform	Hugging Face

A brief comment: Aero-1-Audio used 150 million parameters to prove that small models can surpass large models such as Whisper in audio tasks.

User and market recognition

Gradually build user awareness in the field, and product capabilities are used by content creators and teams to improve work efficiency. Some industry users have incorporated it into their daily workflow. It is recommended to refer to the latest official disclosures for specific user scale and industry adoption rate data.

Cost advantage

C-side/Personal: Completely open source and free, you can download and use it from Hugging Face. You need to prepare your own reasoning context (GPU recommended).

Developers: Free and open source, can be integrated into your own applications. 16 H100 Training 1-day low-cost training regimen that is research team friendly.

Enterprise/Private: Open source allows for self-deployment and fine-tuning. Open licenses such as MIT allow commercial use (the specific license shall be subject to the official).

Main functions

• Long Audio Processing: Supports continuous audio input up to 15 minutes without segmentation, maintaining contextual coherence.
• Speech Recognition (ASR): High-accuracy speech-to-text conversion, suitable for real-time transcription, meeting minutes, and lecture transcription.
• Complex Audio Analysis: Supports analysis and semantic understanding of multiple audio types such as speech, sound effects, music, etc.
• Command-driven tasks: Extract specific information from the audio or perform specific operations based on instructions.
• Lightweight and efficient: Only 150 million parameters, outperforming the larger Whisper and Qwen-2-Audio in multiple benchmarks.

Model and version evolution

Continuous iterative updates, the latest version introduces performance optimization and new features. Historical version information can be viewed on the official release page. There is no complete public version evolution timeline yet. It is recommended to pay attention to the official announcement to understand the rhythm of feature updates.

Technical advantages

Main type judgment: Basic large model/API infrastructure - lightweight open source audio model.

Performance and Throughput: 1 day of training on just 16 H100 GPUs. FLOP utilization in the inference phase is improved from 0.03 to 0.34 through sequence packing technology. Online service indicators such as TTFT and TPM/RPM are not public and are subject to the official Hugging Face page.

Adaptation Boundary: Best at speech recognition (ASR) and long audio understanding tasks. Excellent performance in following voice commands and understanding audio scenes. But it is not good at generative audio tasks such as music generation and speech synthesis (TTS).

Training method: Adopt a dynamic batch processing strategy based on token length to improve computing resource utilization by grouping samples into predefined token length thresholds.

How to use

Entrance	Description
Hugging Face	Visit huggingface.co/lmms-lab/Aero-1-Audio to download the model
Local inference	Load the model through the Transformers library for inference

Product Pricing

The pricing model is subject to the official real-time page. Usually a freemium or subscription system is used, and basic functions can be used for free. Advanced functions or high-frequency use require paid subscriptions, and users are advised to evaluate the optimal solution based on actual usage.

Application scenarios

• Voice Assistant: Provides efficient speech recognition and understanding capabilities for intelligent voice assistants.
• Real-time Transcription: Quickly transcribe voice content into text, suitable for meetings, lectures, etc.
• Audio Archiving and Search: Add content tags to the audio library and support semantic search.
• Agent Listening Module: Gives the agent the ability to understand long speech and supports multiple rounds of dialogue.

Applicable people

• Individual Users: Content creators and knowledge workers who need AI assistance to improve their daily work efficiency.
• Developers: Technical teams who need to integrate AI capabilities into their own products or services through APIs.
• Enterprise: Organizations seeking to deploy AI at scale in their field.

Summary and Outlook

Aero-1-Audio demonstrates the engineering capabilities of solving big problems with small models - 150 million parameters + 50,000 hours of training data, surpassing Whisper on the ASR benchmark. The training efficiency optimization solution (dynamic batch processing + sequence packaging) has reference value for research teams with limited resources.

Unfit Boundary: Does not provide speech synthesis (TTS) capabilities; inference requires GPU resources; the model only supports English; the vocal separation effect in unstructured long audio is not disclosed; as a research model, the stability and support of production context need to be evaluated by yourself.

Purchase Suggestion: The research team can directly download the model for reproduction and fine-tuning. Before deployment in production, it is necessary to verify the ASR accuracy on the target data and evaluate the GPU inference cost. Pay attention to whether LMMs-Lab will release a larger version or a dedicated fine-tuned version in the future.

Related tools: elevenlabs, udio

Version Info

• Aero-1-Audio release ：Based on Qwen-2.5-1.5B, it supports 15 minutes of continuous audio and is optimal for ASR multi-benchmarks. （~2025-05）
• Aero-1-Audio preview ：Early preview version, core audio processing architecture verification. （~2025-03）