Adapting time‐frequency responses to speech and acoustic classification tasks Free

Much past effort, including our own, has gone into the design of high‐resolution time‐frequency representations. However, as found in most data‐trained automatic classification applications, high resolution can significantly degrade performance. This behavior can be understood from the standpoint of complexity: The higher the resolution and complexity of the time‐frequency representation, the larger the size of the training data needed for an accurate classifier. For most applications, where many regions in time and frequency are not salient to the classification task at hand, lowering resolution in these regions should actually improve classifier performance and generality. A new time‐frequency approach, called ‘‘class‐dependent kernels,’’ which selectively smoothes in time and frequency, has been developed. This approach can be summarized by considering the points of a discrete auto‐ambiguity function, which is the two‐dimensional Fourier transform of a time‐frequency representation, as a set of features, which can be ranked in terms of saliency to the particular classification task. Keeping and classifying based on only a small number of these ambiguity plane points corresponds to a flexible and data‐adaptive smoothing of the corresponding time‐frequency plane. This approach will be shown to be useful for a wide range of acoustic and vibration problems including monitoring acoustic emissions in titanium removal and as an adjunct to cepstra in speaker‐independent phoneme recognition.