· · · Repository hosted by TU Delft Library

Stereo audio signal is often modeled as a mixture of instantaneously mixed primary components and uncorrelated ambience components. This paper focuses on the estimation of the primary-to-ambience energy ratio, PAR. This measure is useful for signal decomposition in stereo and multichannel audio coding, format conversion, and spatial audio enhancement. The conventional approaches for the estimation of the ratio are based on the ratio of eigenvalues which requires equal energies of the ambience signals. This often leads to an inaccurate estimate of PAR. An alternative measure is proposed which reduces those estimation errors but requires a priori information about the primary component signal. The performance of the method is demonstrated with synthetic signals and a large collection of stereo audio data.

Tempo estimation is a subject of intensive research in the field of Music Information Retrieval, as many applications demand the automatic induction of the tempo of musical excerpts. In such applicationsit is desired that a correct tempo estimation would be available to the system at about the same time that the tempo is detected by a human listener. This is technically very difficult because the human listeners are able to use higher level context cues to conduct tempo detection. In fact, many algorithms proposed for tempo detection in the past require a long signal segment for reliable results in tempo estimation. This is clearly a problem in contents such as radio programs, where the rhythmic music content may alternate with, for example, speech segments. There is a wide range of literature methods related to the topic of tempo estimation. So far, tempo estimation systems follow a general scheme that consists of two main steps. In the first step, a feature list is created which is used in the second step in order to detect recurrences of certain events in it. Many different approaches have been proposed in the past for the implementationof the above stages. In this thesis, a new approach to the implementation of the first step is proposed, along with the addition of a final step that will enhance the whole tempo estimation procedure. The proposed method for the extraction of the feature list is based ontransient detection. The term transient is used to describe these points in the time representation of the input signal where abrupt changes take place in its amplitude. The detection is conducted using Gammatone subspace analysis and adaptive Linear Prediction Error Filters. The transient detection function produced from this processing is further processed resulting to the necessary feature list. After the second step, where the feature list is fed as an input to a bank of comb filters resonators, the application of a model that approximates the tempo perception by human listeners is proposed. The later will enhance the results of tempo estimation with perceptual information. The evaluation of the proposed system is done using accuracy measures and musical excerpts obtained from the ISMIR 2004 Tempo Induction Evaluation Exchange benchmark corpus, also used from the first ever attempt to conduct systematic comparison of tempo estimation systems. The results of the evaluation indicate that the proposed method compares favourably with other, state-of-the-art tempo estimation methods, using only one frame of the musical excerpts when most of the literature methods demand the processing of the whole piece.