Cleaning and Fine Structuring SNR Samples for Flexible Class labeling

Amplitude modulation spectrogram (AMS) feature is a useful and efficient time frequency representation of speech, in particular for speech analysis and synthesis. However, due to the tedious calculation for signal to noise ratio (SNR) for each sub-band in each frame, data preparation is always a time consuming work. Two major problems need to be addressed here. Firstly, AMS features and associated SNR values are created simultaneously and saved synchronously in order to be accessed conveniently. It will cause memory trouble when sample number is huge. Secondly, SNR samples are used and refreshed ceaselessly in classification. It is always desirable that least operations will be required on existing dataset. The research reported in this paper describes our investigations to these problems with the aim to label SNR values which are associated with AMS feature vectors. Since these labels and associated AMS feature vectors are to be employed as training data in a deep belief network (DBN) for SNR classification, they should possess some essential characteristics, e.g., coverage of the feature space, approximate sample number in each class, easy to extend, etc. Solutions to these considerations are explained one by one, and modifications of dataset in practice are recommended as well.