Joint source-filter optimization for robust glottal source estimation in the presence of shimmer and jitter

We propose a glottal source estimation method robust to shimmer and jitter in the glottal flow. The proposed estimation method is based on a joint source-filter optimization technique. The glottal source is modeled by the Liljencrants–Fant (LF) model and the vocal-tract filter is modeled by an auto-regressive filter, which is common in the source-filter approach to speech production. The optimization estimates the parameters of the LF model, the amplitudes of the glottal flow in each pitch period, and the vocal-tract filter coefficients so that the speech production model best describes the observed speech samples. Experiments with synthetic and real speech data show that the proposed estimation method is robust to different phonation types with varying shimmer and jitter characteristics.