Ambient Noise Classification for Automatic Speaker Identification

This paper proposes two methods for acoustic ambient noises classification. The classification is based on the Kurtosis coefficient and the Bhattacharyya distance. Five colored acoustic noises, some captured in different environments and a White artificially generated, were used to perform the classification methods. These noises were obtained from NOISEX-92 database. Automatic speaker identification experiments were conducted using TIMIT speech database, corrupted with the acoustic noises. Mismatch conditions (SNR of 10 dB, 15 dB and 20 dB) were also examined in the experiments. The performances presented considerable variations among the different acoustic noises. The results show that the noise classification obtained with the proposed methods could detect the differences in the speaker identification accuracies. The MFCC (Mel-Frequency Cepstrum Coefficients) and GMM (Gaussian Mixture Models) were applied for the identification experiments.