Perceptual Audio Coding

1
Perceptual Audio Coding

• By Nick Hone and Colin Carnegie
• July 25, 2005

N
2
Why Perceptual Coding?

• The data rate of a standard audio CD is 172
  KB/sec
• Home Internet connections had speeds of 5 KB/sec
• Perceptual coding exploits the imprecision of the
  human ear

N
3
Perceptual Coding Standards

• Audio
• MPEG-1 Part 3 Layers 1, 2, and 3
• Ogg Vorbis
• Video
• MPEG-1 Part 2, MPEG-2 Part 2
• H.264
• Ogg Theora

N
4
Human Auditory Limitations

• Frequency resolution within critical bands
• Loud signals overpower adjacent signals
• Quantization noise is less audible at these
  points - we can save bits!

C
5
Compression System
C
6
Filter Bank

• 32 band polyphase filter
• Overlapping bands
• Non Reversible
• Linear (unlike 25 critical bands)
• Noise introduced is negligible

N
7
Psychoacoustic Model

• Operates on 12 samples
• Hann weighted FFT/separated into bands
• Tonal/non tonal power separated
• Non tonal information shifted to geometric mean
• Masking function calculated, using global and
  dynamic threshold values

C
8
Bit Allocation

• Quantization required for each subband determined
  with iterative process
• Starting from total bits, distribute a bit to a
  sub-band with the highest signal to mask ratio,
  then recalculate ratios

C
9
Polyphase Filter Implementation

• Sub-band filtering achieved using modified
  discrete cosine transform
• Implemented in MATLAB with loops evaluating the
  following equation

• C is the filter response
• X is the PCM samples
• i denotes the current sub-band

N
10
Bitstream Formatting 1

• MPEG-1 Layer 1 frames are independent
• Frames can be randomly accessed
• Each frame stores 384 PCM samples
• Each frame is identified by a valid header

N
11
Bitstream Formatting 2

• Header contains data such as sampling rate,
  presence of CRC field
• The infrequently used 16 bit CRC field allows
  error checking
• All sub-bands have a bit allocation field
• Sub-bands with zero bits do not have scale factor
  or sample fields

N
12
Analysis Frames
N
13
Implementation-Bit Allocation

• Bit allocation and psychoacoustic models are
  combined
• Applied to the 32 sub-bands instead of critical
  bands

C
14
Implementation-Bit Allocation

• Average power calculated for each sub-band
• Sub-band power scaled by frequency response of
  humans ear sensitivity
• Sub-bands adjacent to high power sub-bands are
  reduced.
• Iterative bit allocation process distributes bits
  to highest power sub bands

C
15
Power in a Signal
C
16
Bit Allocation
C
17
Test Signal Spectrogram
N
18
Test Signal

• Spectrum spans 0 to 20 kHz
• Most power near 0 Hz
• Noise is relatively low (blue-green)

N
19
Compressed Signal Spectrogram
N
20
Compressed Signal

• Spectrum spans 0 to 4 kHz
• Almost all power concentrated in 0 to 4 kHz
• High frequencies replaced with noise

N
21
Conclusion

• MPEG is really, really hard
• Concepts are simple to prove
• The devil is in the details

C