Abstract: Language identification is heavily influenced by the real noise environment, resulting in poor identification results. To solve this problem, an image processing method for language identification is proposed based on the logarithmic gray-scale speech spectrogram. The logarithmic gray-scale speech spectrogram is obtained by combining the human auditory properties and the voice filtered in real noise environments according to the different distribution of noise and speech on the speech spectrogram. Then, the principal component of the spectrogram is extracted as language features and, a residual neural network model is used for training and testing. Experimental results show that the average identification rate of the proposed method is improved by 27.5% in the noisy cockpit of a Blackburn Buccaneer compared to the linear grey-scale speech spectrogram method. In other noisy environments, the average identification rate is also improved.

Key words: language identification, real noise environment, logarithmic gray-scale speech spectrogram, residual neural network, image processing