Aim: Heart murmur detection can provide a preliminary diagnosis of heart disease, and has become increasingly important in assisting clinical diagnosis and treatment in recent years. The purpose of this study is to construct an automatic detection system for heart murmurs. Methods: We build a learnable filter-based transformer architecture. The learnable filter is embedded between the embedding layer and the encoder layer of the transformer. The parameters of the filter are optimized by Adam to adaptively represent any filter in the frequency domain, thereby achieving the effect of adaptive noise reduction. Then, the transformer encoder module captures the long-term dependencies of the heart sound signal, allowing the network to learn more effective features from the input signal. Finally, the final classification result will be obtained according to the voting rules we set. Results: Our (Bear _FH) method is trained and validated on public datasets proposed by the challenge. In the formal phase of the challenge, testing the trained algorithm with a hidden test set, we achieved challenge metric scores (weight accuracy and cost) of 0.402 and 24406 on the murmur detection task. Our final ranking is 39th. We achieved a challenge metric scores (weight accuracy and cost) of 0.5 and 15982 on the clinical outcome identification task. Our final ranking is 35th.