The transmission gear is accompanied by a large amount of noise during operation, which affects the accuracy of fault diagnosis based on sound signals. Therefore, in response to the problem of difficult feature extraction and low accuracy of fault diagnosis based on sound signals of transmission gears, this paper proposes a transmission fault diagnosis method based on sound signals. Firstly, combining wavelet threshold denoising and CEEMDAN denoising methods is used to denoise the collected sound signals, improving the signal-to-noise ratio of the sound signals. Then, the Mel frequency cepstral coefficient features of the sound signals are extracted to construct a feature vector dataset. Finally, the CNN-BiLSTM network model is used to train and test the feature vectors, achieving fault diagnosis of transmission gears based on sound signals. The sound signals of different faults collected from a gear fatigue test bench with a center distance of 91.5 mm are trained and tested, and the accuracy of fault diagnosis in the test dataset reaches 94.4%, effectively verifying the high reliability of the proposed gear transmission fault diagnosis method based on sound signals.