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 t