This paper proposes a new variant of the least square autoregressive (LSAR) method for speech reconstruction, which can estimate
via least squares a segment of missing samples by applying the linear
prediction (LP) model of speech. First, we show that the use of a single
high-order linear predictor can provide better results than the classic
LSAR techniques based on short- and long-term predictors without the
need of a pitch detector. However, this high-order predictor may reduce
the reconstruction performance due to estimation errors, especially in the
case of short pitch periods, and non-stationarity. In order to overcome
these problems, we propose the use of a sparse linear predictor which
resembles the classical speech model, based on short- and long-term correlations, where many LP coefficients are zero. The experimental results
show the superiority of the proposed approach in both signal to noise
ratio and perceptual performance.