Speech/Music Discrimination using Spectral Peak Feature for Speaker Indexing

We present a new speech/music discrimination method based on spectral peak feature and spectral peak's duration threshold. The focus is feature extraction that reflects the spectral peak's duration characteristic. Also, we consider fast discrimination and high performance. We extract the spectral peak feature from audio spectrum's each peak track and normalize the feature by length of segment. The extracted spectral peak's duration feature can be easily discriminated the speech and music using the duration threshold. We evaluate our method on speech (Korean, English, Chinese and Japanese) and various kinds of pop-music (ballad, rock etc.) for 26,773 seconds of audio data. The average accuracy is 96.21% for speech and 89.49% for music. It was found from the experimental result that our feature vector is suitable for speech/music discrimination and it is computational efficient