Abstract:Recently, reinforcement learning (RL) has been shown to greatly enhance the reasoning capabilities of large language models (LLMs), and RL-based approaches have been progressively applied to visual multimodal tasks. However, the audio modality has largely been overlooked in these developments. Thus, we conduct a series of RL explorations in audio understanding and reasoning, specifically focusing on the audio question answering (AQA) task. We leverage the group relative policy optimization (GRPO) algorithm to Qwen2-Audio-7B-Instruct, and our experiments demonstrated state-of-the-art performance on the MMAU Test-mini benchmark, achieving an accuracy rate of 64.5%. The main findings in this technical report are as follows: 1) The GRPO algorithm can be effectively applied to large audio language models (LALMs), even when the model has only 8.2B parameters; 2) With only 38k post-training samples, RL significantly outperforms supervised fine-tuning (SFT), indicating that RL-based approaches can be effective without large datasets; 3) The explicit reasoning process has not shown significant benefits for AQA tasks, and how to efficiently utilize deep thinking remains an open question for further research; 4) LALMs still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further exploration. Our project is available at https://github.com/xiaomi-research/r1-aqa and https://huggingface.co/mispeech/r1-aqa.
Abstract:This challenge aims to evaluate the capabilities of audio encoders, especially in the context of multi-task learning and real-world applications. Participants are invited to submit pre-trained audio encoders that map raw waveforms to continuous embeddings. These encoders will be tested across diverse tasks including speech, environmental sounds, and music, with a focus on real-world usability. The challenge features two tracks: Track A for parameterized evaluation, and Track B for parameter-free evaluation. This challenge provides a platform for evaluating and advancing the state-of-the-art in audio encoder design.
Abstract:Researchers have reported high decoding accuracy (>95%) using non-invasive Electroencephalogram (EEG) signals for brain-computer interface (BCI) decoding tasks like image decoding, emotion recognition, auditory spatial attention detection, etc. Since these EEG data were usually collected with well-designed paradigms in labs, the reliability and robustness of the corresponding decoding methods were doubted by some researchers, and they argued that such decoding accuracy was overestimated due to the inherent temporal autocorrelation of EEG signals. However, the coupling between the stimulus-driven neural responses and the EEG temporal autocorrelations makes it difficult to confirm whether this overestimation exists in truth. Furthermore, the underlying pitfalls behind overestimated decoding accuracy have not been fully explained due to a lack of appropriate formulation. In this work, we formulate the pitfall in various EEG decoding tasks in a unified framework. EEG data were recorded from watermelons to remove stimulus-driven neural responses. Labels were assigned to continuous EEG according to the experimental design for EEG recording of several typical datasets, and then the decoding methods were conducted. The results showed the label can be successfully decoded as long as continuous EEG data with the same label were split into training and test sets. Further analysis indicated that high accuracy of various BCI decoding tasks could be achieved by associating labels with EEG intrinsic temporal autocorrelation features. These results underscore the importance of choosing the right experimental designs and data splits in BCI decoding tasks to prevent inflated accuracies due to EEG temporal autocorrelation.
Abstract:Speech emotion recognition (SER) is to study the formation and change of speaker's emotional state from the speech signal perspective, so as to make the interaction between human and computer more intelligent. SER is a challenging task that has encountered the problem of less training data and low prediction accuracy. Here we propose a data augmentation algorithm based on the imaging principle of the retina and convex lens, to acquire the different sizes of spectrogram and increase the amount of training data by changing the distance between the spectrogram and the convex lens. Meanwhile, with the help of deep learning to get the high-level features, we propose the Deep Retinal Convolution Neural Networks (DRCNNs) for SER and achieve the average accuracy over 99%. The experimental results indicate that DRCNNs outperforms the previous studies in terms of both the number of emotions and the accuracy of recognition. Predictably, our results will dramatically improve human-computer interaction.