1. Jason A. Hockman McGill University 24 January 2008
McGill University >Schulich School of Music > Music Technology > MUMT 611
FLAC
free lossless audio codec
Jason A. Hockman
McGill University
24 January 2008

2. Presentation Overview
What is FLAC?
Codec Format
Lossless Codec Comparison
FLAC
free lossless audio codec

3. Presentation Overview
What is FLAC?
Codec Format
Lossless Codec Comparison
FLAC
free lossless audio codec

4. Presentation Overview
What is FLAC?
Codec Format
Lossless Codec Comparison
FLAC
free lossless audio codec

5. FLAC IS…
the Free Lossless Audio Codec (chief designer: Josh Coalson)
free: no cost / format is fully public / no patent
lossless: PCM data incurs no distortion (verified per frame)
fast: fast decode due to asymmetry
seekable: sample-accurate seeking
searchable: independent frames contain sync code and CRCs
decoders can pick up mid-stream, and errors are limited to local frame
cyclic redundance check
FLAC
free lossless audio codec

6. FLAC Codec Format WAV or AIFF containers for RAW PCM audio
FLAC (Apple Lossless) use PCM, but use compression to cut file size %
FLAC
free lossless audio codec

7. FLAC Codec Format FLAC format supports: 1 - 8 channels
bit (currently )
all standard SR <=192 kHz
FLAC
free lossless audio codec

8. FLAC Codec Format architecture free lossless audio codec 1 2 3 4
interchannel
decorrelation
residual
coding
blocking
LPC
Blocking - input is separated into sequential blocks
Interchannel Decorrelation - encoder exploits correlation between channels
Prediction and Residual Coding fall under the larger header of linear predictive coding(LPC) methods.
FLAC
free lossless audio codec

9. FLAC Blocking tradeoff between required data space for
headers and prediction accuracy
Ideally block size will be large enough to not allocate too much data space to headers, but not to the point where data is negatively affected by variation in data stream
FLAC
free lossless audio codec

10. FLAC Blocking tradeoff between required data space for
headers and prediction accuracy
16 samples(min) samples(max)
Ideally block size will be large enough to not allocate too much data space to headers, but not to the point where data is negatively affected by variation in data stream
FLAC
free lossless audio codec

11. FLAC Blocking tradeoff between required data space for
headers and prediction accuracy
16 samples(min) samples(max)
Ideally block size will be large enough to not allocate too much data space to headers, but not to the point where data is negatively affected by variation in data stream
currently block size determined by sample rate
FLAC
free lossless audio codec

12. FLAC Blocking for each channel free lossless audio codec block header
subframe
for each channel
Before prediction is preformed, each block is prepared one channel at a time, and is coded as a subframe, along with a header providing information about how the audio was coded
FLAC
free lossless audio codec

13. Interchannel Decorrelation
header A C C B A
channel 2
header A B B B A
on a frame-by-frame basis, channels are coded
independently
mid-side
left-side
right-side
Here the characters A B and C are simply used to define similarity between channels
Each channel is coded separately
FLAC
free lossless audio codec

14. Interchannel Decorrelation
header A C C B A
channel 2
header A B B B A
on a frame-by-frame basis, channels are coded
independently
mid-side
left-side
right-side
Here the characters A B and C are simply used to define similarity between channels
The average is taken and called the mid, and the residual of each side is coded
FLAC
free lossless audio codec

15. Interchannel Decorrelation
header A C C B A
channel 2
header A B B B A
on a frame-by-frame basis, channels are coded
independently
mid-side
left-side
right-side
Here the characters A B and C are simply used to define similarity between channels
Left side is coded and the difference given by LEFT MINUS RIGHT is coded as the side
FLAC
free lossless audio codec

16. Interchannel Decorrelation
header A C C B A
channel 2
header A B B B A
on a frame-by-frame basis, channels are coded
independently
mid-side
left-side
right-side
Here the characters A B and C are simply used to define similarity between channels
Right side is fully coded and the difference given by RIGHT MINUS LEFT
FLAC
free lossless audio codec

17. Prediction
Approximation
Residual
polynomial fitting
LPC
difference
verbatim
constant
FLAC approximates signal, then subtracts this approximation from the original signal to create the residual
2 methods of approximation:
polynomial fitting - which is much faster but not as accurate
LPC
Fixed LP
FIR LP
FLAC
free lossless audio codec

18. Prediction
Approximation
Residual
polynomial fitting
LPC
difference
verbatim
constant
FLAC approximates signal, then subtracts this approximation from the original signal to create the residual
4 modes of modeling available to lpc.
Verbatim: residual is = signal, as x-hat is understood to be zero
Fixed LP
FIR LP
FLAC
free lossless audio codec

19. Prediction
Approximation
Residual
polynomial fitting
LPC
difference
verbatim
constant
FLAC approximates signal, then subtracts this approximation from the original signal to create the residual
constant
Fixed LP
FIR LP
FLAC
free lossless audio codec

20. Prediction
Approximation
Residual
polynomial fitting
LPC
difference
verbatim
constant
FLAC approximates signal, then subtracts this approximation from the original signal to create the residual
2 methods of approximation:
polynomial fitting - which is much faster but not as accurate
LPC
Fixed LP
FIR LP
FLAC
free lossless audio codec

21. Prediction
Approximation
Residual
polynomial fitting
LPC
difference
verbatim
constant
FLAC approximates signal, then subtracts this approximation from the original signal to create the residual
2 methods of approximation:
polynomial fitting - which is much faster but not as accurate
LPC
Fixed LP
FIR LP
FLAC
free lossless audio codec

22. Linear Prediction Coding
excitation
synthesis filter
LPC is a source-filter analysis-synthesis method to approximate sound generation by modeling a pulse train or noise signal through an all-pole resonant filter
In LPC the input sample x(n) is approximated by a linear combination of past samples of the input signal, where a is the prediction coefficient associated with the particular sample of x, and p is the order number.
sound
FLAC
free lossless audio codec

23. Linear Prediction Coding
excitation
synthesis filter
residual:
LPC is a source-filter analysis-synthesis method to approximate soound generation by modeling a pulse train or noise signal through an all-pole resonant filter
In LPC the input sample x(n) is approximated by a linear combination of past samples of the input signal, where a is the prediction coefficient associated with the particular sample of x, and p is the order number.
A = coefficients
P = order number
sound
FLAC
free lossless audio codec

24. Linear Prediction Coding
FIR Filter + - + + +
FLAC
free lossless audio codec

25. Linear Prediction Coding
FIR Filter + - + + +
A are prediction coefficients
Z-transform
FLAC
free lossless audio codec

26. Linear Prediction Coding
FIR Filter + - + + +
Z-transform
FLAC
free lossless audio codec

27. Linear Prediction Coding
estimating LPC parameters
residual energy,
Where alpha-k are values that minimize E
FLAC
free lossless audio codec

28. Linear Prediction Coding
estimating LPC parameters
residual energy,
E is minimized,
for
(E)nergy = Sum of squared error across the samples within the frame
Solving the partial derivative of energy function E with respect to variable a-i = 0 for i = 1 to model order p. gives us:
FLAC
free lossless audio codec

29. Linear Prediction Coding
estimating LPC parameters
residual energy,
E is minimized,
for
Solving the partial derivative of energy function E with respect to variable a-i = 0 for i = 1 to model order p. gives us:
Where alpha-k are values that minimize E
yielding:
FLAC
free lossless audio codec

30. Linear Prediction Coding
estimating LPC parameters
substitution:
FEE
EVEN FUNCTION!
Resolved through the levitson-durbin recursion
FLAC
free lossless audio codec

31. Linear Prediction Coding
estimating LPC parameters
substitution:
FEE
EVEN FUNCTION!
Resolved through the levitson-durbin recursion
FLAC
free lossless audio codec

32. Linear Prediction Coding
estimating LPC parameters
substitution:
ACF:
FEE
EVEN FUNCTION!
Resolved through the levitson-durbin recursion
FLAC
free lossless audio codec

33. Linear Prediction Coding
estimating LPC parameters
substitution:
ACF:
FEE
EVEN FUNCTION!
Resolved through the levitson-durbin recursion
follows:
FLAC
free lossless audio codec

34. Linear Prediction Coding
estimating LPC parameters
substitution:
ACF:
FEE
EVEN FUNCTION!
Resolved through the levitson-durbin recursion
follows:
FLAC
free lossless audio codec

35. Linear Prediction Coding
estimating LPC parameters
optimal coefficients minimize residual energy
higher order provides closer approximation
tradeoff between spectral accuracy and timely output
Where alpha-k are values that minimize E
(Zolzer et. al 2002)
FLAC
free lossless audio codec

36. Comparison
FLAC and WavPack are only lossless compression codec that are
open-source
available via no license cost
is streamable
is seekable
Where alpha-k are values that minimize E
FLAC
free lossless audio codec

37. FLAC Comparison free lossless audio codec
compression versus encode speed
compression versus decode speed
(Coalson, 2007)
FLAC
free lossless audio codec

38. FLAC Comparison free lossless audio codec comp versus encode speed
comp versus decode speed
( 2007)
FLAC
free lossless audio codec

39. References
Ariakis, D. WavPack, 2007, (23 January 2008)
Coalson, J. FLAC. Free Lossless Audio Codec, 2007,
(22 January 2008).
Keiler, F., D. Arfib, and U. Zolzer Efficient linear prediction for digital audio
effects. In Proc. DAFX-00 Conference on Digital Audio Effects, Verona.
December 2000.
Zolzer, U DAFx - Digital audio effects. New York: John Wiley and Sons
Where alpha-k are values that minimize E
FLAC
free lossless audio codec