Glottal/Supraglottal Source Separation in Fricatives Based on Non-Stationnary Signal Subspace Estimation

Benjamin Elie (1) , Gilles Chardon (1)

1 L2S - Laboratoire des signaux et systèmes

—The X-GLOS (EXtraction of GLOttal Sources) method for separating the glottal and the supraglottal sources in speech signals is presented in this article. Unlike other periodic/aperiodic decomposition methods that use stationary models of the signal within frames, X-GLOS considers locally varying instantaneous fundamental frequency. Applications on numerically synthesized fricative signals prove the locally non-stationary model to be more robust to moderate and high jitter values than stationary models. A peak picking selection also allows X-GLOS to be less sensitive to high colored noise levels. The gain of performance, in comparison with the reference existing method, is about a couple of dozens of dB in high noise-to-harmonics ratios. X-GLOS can then be specifically used to study the behavior of the voicing and frication noise sources independently, even at vowel-consonants behavior where the voiced source is less powerful that the frication noise in the recorded mixture speech signal.

Mots clés: Periodic/aperiodic decomposition; Speech production; Fricative production

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Benjamin Elie, Gilles Chardon. Glottal/Supraglottal Source Separation in Fricatives Based on Non-Stationnary Signal Subspace Estimation. 2018. ⟨hal-01764890⟩