This paper presents a new approach to elimination of impulsivedisturbances from archive speech signals. The proposedsparse autoregressive (SAR) signal representation is given ina factorized form - the model is a cascade of the so-called formantfilter and pitch filter. Such a technique has been widelyused in code-excited linear prediction (CELP) systems, as itguarantees model stability. After detection of noise pulses usinglinear prediction, the factorized model is converted intoa generic sparse form in order to perform a projection-basedsignal interpolation. It is shown that the proposed algorithmis able to deal favorably with speech signals with strong glottalactivity, which is a serious problem for algorithms basedon the classical AR modeling.
Authors
Additional information
- Category
- Aktywność konferencyjna
- Type
- materiały konferencyjne indeksowane w Web of Science
- Language
- angielski
- Publication year
- 2012
