A fraud identification method and device which combine audio analysis and video analysis and a storage medium. The method comprises the following steps: cutting an audio/video sample to obtain audio/video clips, and assigning a fraud label for each audio/video clip (S10); decoding and preprocessing each audio/video clip to obtain an audio clip and a video clip for each audio/video clip (S20); extracting a voice feature and an emotional expression feature from each audio clip and video clip respectively (S30); training a support vector machine by using the voice feature of each audio clip and the emotional expression feature of each video clip in combination with the fraud label so as to obtain a voice analysis model and an emotional expression analysis model (S40); acquiring audio/video data of an object to be identified (S50); extracting a voice feature and an emotional expression feature of the audio/video data (S60); inputting the voice features and emotional expression features into the voice analysis model and emotional expression analysis model respectively, and outputting fraud probabilities P1 and P2 of the object to be identified (S70); and calculating the weight of P1 and P2 to obtain a fraud identification result of the object to be identified (S80).