With the development of deep fake technology, synthetic speech is created easier by forgery techniques based on text-to-speech and voice conversion, which poses a challenge to automatic speaker verification systems. Existing methods demonstrate excellent performance on public databases, but most methods are weak in detecting compressed speech commonly used in social networks, such as MP3 and AAC. We believe that if the classifier has compressed information as a priori knowledge, it will help the classifier make a more accurate decision when detecting compressed speech. To solve this issue, a multi-branch residual network with a compression feature embedding module is proposed in this paper. The feature embedding module is used to integrate the authenticity feature and compression feature. Our method is evaluated on the ASVspoof database and experimental results show the effectiveness of the proposed method for detecting compressed speech.