Hi everyone,

Thanks for your questions! Let me clarify the resource restrictions for the challenge:

1. CPU and Training Restrictions:Inference: Yes, the CPU restriction (e.g., ARM Cortex-A53) applies only to the inference stage. This means that during deployment on edge devices, the model should run efficiently on a single CPU with no GPU or TPU support. Training: You are allowed to use GPUs during the training phase. The 6-hour training time limit applies regardless of the hardware used for training.
2. Pretrained Models:By "not allowing pretrained models," we mean that you cannot use any models that have been pretrained on external datasets. This includes models like those from Hugging Face or OpenAI's Whisper. The goal is to encourage the development of models trained from scratch using the provided dataset only. Fine-tuning of public models is also not allowed for this challenge. All participants must build their models from the ground up.
3. Inference Time:The statement "Inference time needs to be 2 minutes or less" refers to the total time it takes for the model to make predictions on the entire test dataset (2,000 audio clips) during inference. This is intended to simulate the real-time constraint of a ~50ms inference time per audio clip on edge devices. Therefore, the total inference process should be completed in under 2 minutes.
4. Feature Engineering:Yes, the inference time includes any feature engineering or preprocessing steps required to make predictions. The goal is to ensure that the entire process, from receiving raw audio to outputting predictions, is optimized for real-time deployment on edge devices.

I hope this clears things up! If you have any more questions, feel free to ask. Good luck to everyone, and we look forward to seeing your solutions on the leaderboard!