SIGNAL zyxwv PROCESSING ELSEVIER Signal Processing 53 (1996) 35-45 Complex AM signal model for non-stationary signals Pradip Sircar *, Mohanjeet Singh Syali zyxwvutsrqponmlkjihgfedcbaZYX Department of’ Electrical Engineering, Indian Institute qf Technology Kanpur, Kanpur 208 016, India Received 17 June 1994; revised 22 February 1996 Abstract A novel signal model consisting of a weighted sum of complex amplitude modulated signals is proposed as a suitable representation for non-stationary signals like speech. The estimation of model parameters is carried out bq utilizing the accumulated autocorrelation functions of the modelled signal. The developed model is first fitted on noise-corrupted synthesized data, and then on sampled voiced speech data. The study demonstrates the suitability of the model. Zusammenfassung Ein neues Signal Modell, bestehend aus einer gewichteten Summe von komplexen amplitudenmodulierten Signalen, wird als eine geeignete Reprisentation fiir nichtstationgre Signale, wie z.B. Sprache, vorgeschlagen. Die Schgtzung der Modellparameter wird unter Benutzung der akkumulierten Autokorrelationsfunktionen des modellierten Signals vorgenommen. Das entwickelte Model1 wird zunlchst an verrauschte synthetisierte Daten angepaRt und danach an abgetastete stimmhafte Sprachdaten. Die Studie demonstriert die Eignung des Modells. Rksumi! Nous proposons un modile de signal nouveau consistant en une somme de signaux modulCs en amplitude complexes pour la reprksentation de signaux non stationnaires tels que la parole. L’estimation des parambtres du modkle est optrke en utilisant les fonctions d’autocorrtlation accumulCes du signal mod6lisC. Le mod&e dCvelopp6 est tout d’abord appliqut: B des donnies bruit&es synthktiques puis & des donnkes de parole voistes 6chantillonn6es. Cette ttude dkmontre les qualitits du modile. Ke,vwwds: Parametric modelling; Non-stationary signals; Time-varying model; Complex AM signal model 1. Introduction Parametric modelling of signal plays an impor- tant role in signal analysis, signal synthesis and various other signal processing applications. Satis- factory performance of parametric signal modelling depends crucially on the suitability of selected model, and then on the accuracy of estimation of model parameters [S]. Time series or autoregressive moving-average (ARMA) model is extensively used in the problem -__ of signal representation. The modern approach of *Corresponding author. E-mail: sircar@iitk.ernet.in. spectral estimation is mainly based on the use of 0165- 1684,/96~$15.00 S 1996 Elsevier Science B.V. All rights reserved PII SOl65-1684(96)00074-6