FLAC uses four methods for modeling the input signal:
Verbatim. This is essentially a zero-order predictor of the signal. The predicted signal is zero, meaning the residual is the signal itself, and the compression is zero. This is the baseline against which the other predictors are measured. If you feed random data to the encoder, the verbatim predictor will probably be used for every subblock. Since the raw signal is not actually passed through the residual coding stage (it is added to the stream 'verbatim'), the encoding results will not be the same as a zero-order linear predictor.
Constant. This predictor is used whenever the subblock is pure DC ("digital silence"), i.e. a constant value throughout. The signal is run-length encoded and added to the stream.
Fixed linear predictor. FLAC uses a class of computationally-efficient fixed linear predictors (for a good description, see audiopak and shorten). FLAC adds a fourth-order predictor to the zero-to-third-order predictors used by Shorten. Since the predictors are fixed, the predictor order is the only parameter that needs to be stored in the compressed stream. The error signal is then passed to the residual coder.
FIR Linear prediction. For more accurate modeling (at a cost of slower encoding), FLAC supports up to 32nd order FIR linear prediction (again, for information on linear prediction, see audiopak and shorten). The reference encoder uses the Levinson-Durbin method for calculating the LPC coefficients from the autocorrelation coefficients, and the coefficients are quantized before computing the residual. Whereas encoders such as Shorten used a fixed quantization for the entire input, FLAC allows the quantized coefficient precision to vary from subframe to subframe. The FLAC reference encoder estimates the optimal precision to use based on the block size and dynamic range of the original signal.
Så den anpassar och använder olika metoder beroende på hur indatat ser ut, fast den kan ändra metoder flera gånger i samma fil. Väldigt fiffigt faktiskt. Jag gillar det med "linear predictor", man gör en förstörande snålkodning med en approximation. Sen kodar man "the error signal", alltså man sparar även felet som blir mellan den riktiga signalen och den approximerade signalen. Då kan man sen återskapa ursprunget. Om jag nu förstått det hela rätt.
