Abstract: To address the problem that features unrelated to language identification in broadcast audio have an impact on the language 
identification results, an improved language identification method based on gamma frequency cepstrum coefficients with improved feature parameters is proposed. By extracting the energy spectral envelope of each frame, the speaker-related features are removed, filtered by a Gammatone filter banks, and then by the discrete cosine transform and cepstrum lifting to obtain the improved gamma frequency cepstrum feature parameters. The feature parameters extracted from broadcast audio signal were input into hidden Markov model for training and testing， and the language identification results were obtained. The results show that the proposed method can effectively improve the language identification accuracy for broadcast audio, which is better than the currently used gamma frequency cepstrum coefficient features and their derivatives.

Key words: broadcast audio language identificaition, energy spectrum envelope, cepstrum lifting, improved gamma frequency cepstrum coefficient