In many applications of audio signal processing modeling of the signal is required. The most commonly used approach for audio signal modeling is to assume the audio signal as an (autoregressive) AR-process where the audio signal is locally stationary over a relatively short time interval. In this case the audio signal can be modeled with an all-pole IIR (infinite impulse response) filter, which leads to LPC (linear predictive coding) where the current input sample is predicted by a linear combination of past samples of the input signal. However, in practice the relatively short time interval (i.e. a frame) where the signal is stationary will vary significantly in the audio signal data stream. Also the information content of the frames will show considerable variation. For a proper modeling of an audio signal it is essential that a suitable frame size and appropriate number of model parameters is used instead of a constant frame size and model order. In this paper we present an adaptive frame-by-frame technique for modeling audio signals, which automatically adjusts the optimal modeling frame size and the optimal number of model parameters for each frame.