雷军：第二届音频编码器能力挑战赛明年9月将同步亮相Interspeech 2026，已开放报名

Core Insights - Xiaomi, in collaboration with Surrey University, Tsinghua University, and Haitian Ruisheng, has launched the second Audio Encoder Capability Challenge (AECC), which will be showcased at the Interspeech 2026 conference in September 2026, with registration now open [1][3][14] Challenge Overview - The challenge aims to enhance audio encoders for large audio language models (LALMs), addressing the current reliance on a single technology, specifically the OpenAI Whisper Encoder, which limits diversity in model architecture and overall capabilities [3][14] - The competition will focus on evaluating the understanding and feature representation capabilities of audio encoders in complex real-world scenarios [3][14] Evaluation Methodology - Participants are required to submit pre-trained encoder models, while the training and evaluation of downstream tasks will be conducted by the organizers using the open-source evaluation system XARES-LLM [5][15] - The XARES-LLM system will automatically download training data, train models, and test various downstream tasks, providing scores for each task [5][15] Training Data - Unlike most competitions, this challenge emphasizes both model design and data utilization, allowing participants to use any publicly accessible data for training [18] - A supplementary dataset provided by Haitian Ruisheng includes various environmental noise samples from eight commercial datasets, covering diverse real-world scenarios [6][18] Competition Tracks - Two tracks are established: Track A focuses on traditional classification tasks, while Track B emphasizes understanding and expression capabilities [19][21] - All submissions will be evaluated across both tracks, with independent rankings for each [19][21] Registration and Submission - Participants must complete the registration by January 25, 2026, and submit their encoder code and model files by February 12, 2026 [22][23] - A technical report must be submitted by February 25, 2026, which can also be submitted as a conference paper [23]