Improving a Time-Stretching Technique based on Harmonic/Percussive Separation

Transient smearing, transient duplication and transient skipping are well-known artefacts introduced by time-stretching (TS) techniques. Aiming to avoid these artefacts transient preservation extensions are applied to TS techniques. Typically, the transients are detected in time/frequency regions. Then, a transient preservation processing is applied to these regions. The TS Technique based on Harmonic/Percussive Separation (HPS) relies on a conceptually easy approach to transient preservation that does not need an explicit transients detection in time. This approach is based on the separation of the input signal into the Harmonic and the Percussive components, where ideally only and exclusively the latter contains transients. A phase vocoder is applied to the harmonic component, whereas an overlap-add with parameters optimized to process transients is applied to the percussive component. The authors of the technique claimed that it preserves effectively transients and they highlighted the simplicity of their approach in comparison to other methods based on explicit transients detection in time. In this thesis, we describe and criticize the mentioned technique. We then propose to change the HPS method used to compute the components and evaluate the audio quality of the transformations obtained using the two systems. We found out that the proposed system improves the original on complex harmonic/percussive mixtures. However, the leaking of harmonic elements into the percussive component and the inverse leaking are often cause of not negligible TS artefacts in both systems, such as pitch distortion and transient smearing. We conclude suggesting that the technique could benefit from the use of a multi-resolutional phase vocoder avoiding the overlap-add. Implementing a scheme for explicit transient detection might be necessary