Aero-1-Audio - A lightweight audio model launched by LMMs-Lab

What is Aero-1 Audio?

Aero-1-Audio is a lightweight audio model developed by LMMs-Lab, built on Qwen-2.5-1.5B, containing only 150 million parameters. Designed specifically for long audio processing, it supports continuous audio input up to 15 minutes without segmentation, while maintaining contextual coherence. Aero-1-Audio performs exceptionally well in Automatic Speech Recognition (ASR) tasks, achieving high accuracy, and excels in complex audio analysis and instruction-driven tasks.

Main functions of Aero-1 Audio

• Long audio processingCapable of processing long-term data. 15 minutes It provides continuous audio without the need for segmentation, maintaining contextual coherence, and is particularly suitable for processing long audio content.
• Automatic Speech Recognition (ASR)It performs exceptionally well in speech recognition tasks, accurately converting speech into text, and is suitable for scenarios such as real-time transcription, meeting minutes, and lecture transcription.
• Complex audio analysisIt supports the analysis of various audio types such as speech, sound effects, and music, and can understand the semantics and emotions in audio, making it suitable for the classification and analysis of audio content.
• Instruction-driven tasksIt supports command-driven audio processing tasks, such as extracting specific information from audio or performing specific operations based on commands, and is suitable for applications such as smart voice assistants.

The technical principles of Aero-1-Audio

• Lightweight design and high performanceThe Aero-1-Audio, with only 150 million parameters, is relatively small in scale but performs exceptionally well in multiple audio benchmarks, outperforming larger models such as Whisper and Qwen-2-Audio.
• Efficient training methodsAero-1-Audio has a relatively small training dataset, using only about 5 billion tokens (equivalent to 50,000 hours of audio), far fewer than other large models. Through high-quality filtered data and optimized training strategies, training can be completed in one day, requiring only 16 H100 GPUs.
• Dynamic batch processing and sequence packaging technologyAero-1-Audio employs a dynamic batching strategy based on token length, significantly improving computational resource utilization by grouping samples into predefined token length thresholds. Through sequence packing technology combined with Liger kernel fusion, the model's FLOP utilization increased from 0.03 to 0.34, further enhancing training efficiency.
• Multitasking capabilityAero-1-Audio excels in speech recognition (ASR) tasks, demonstrating powerful capabilities across multiple dimensions, including audio analysis and understanding, voice command following, and audio scene understanding. For example, it achieves the lowest word error rate (WER) on the AMI, LibriSpeech, and SPGISpeech datasets.

Aero-1-Audio's project address

• HuggingFace model library:https://huggingface.co/lmms-lab/Aero-1-Audio

Application scenarios of Aero-1-Audio

• voice assistantIt provides efficient speech recognition and understanding capabilities for intelligent voice assistants.
• Real-time transcriptionIt can quickly transcribe speech into text, making it suitable for meetings, lectures, and other similar events.
• Archiving UnderstandingAdd content tags to the audio library to support semantic search.
• Hearing moduleIt enables intelligent agents to understand long speech and supports multi-turn dialogues.