A state-space model with neural-network prediction for recovering vocal tract resonances in fluent speech from Mel-cepstral coefficients