Audio Compression

1
Audio Compression

• By Eric Wietgrefe

2
Audio Compression

• Lossy
• Perceptual
• .mp3
• Lossless
• Monkey

3
MPEG-1

• Perceptual, not lossless coding
• Auditory Masking
• Has 3 layers
• Standardizes the output, not the encoder
• Allows all types of sound waves
• Used in DVDs, DirecTV, Internet Radio

4
MPEG-1 Layers

• Three different coding schemes
• They differ in complexity
• Layer III
• This layer is where .mp3 comes from

5
.mp3

• Is a codec based on perceptual coding
• Yields low bit-rate without much loss of quality
• No specific source model

6
Lossless Compression

• Ratio at best of 31, while lossy at best 121
• Lossless has higher quality
• Compression of an audio signal
• Audio Signal ? Framing ? xn ? Intrachannel
  Decorrelation ? en ? Entropy Coding ?
  Compressed Signal
• Reached the limit of compression

7
Analysis

• SNR
• FFT
• Framing, Segmentation, Classification

8
SNR

• ABSTRACT
• Signal is what you want to hear
• Performer, instruments, etc.
• Noise is what you do not want
• Static, noise over our hearing range, etc.
• Novice and Expert Example

9
Noise

• Easy to ignore data that is not there
• Easy to mask what is needed with data that is
  irrelevant
• Cause and effect relationships

10
FFT

• Filter testing was based on the three variables.
• 2 approaches
• One is to apply time-frequency smearing only to
  the reference signal and use it as a masking
  threshold for the error signal.
• The other is to use the complete perceptual model
  on both the coded signal and the reference signal
  and compare the internal representations of both
  signals.

11
Segmentation Classification

• Segmentation is taking audio signal and
  partitioning it
• Framing
• Classification is determining what made the
  sound.
• Like classification in biology
• Gaussian
• K-Nearest Neighbors

12
Intrachannel Decorrelation

• Used with framing
• Remove redundancy
• Linear predictor
• Xn Xn-1
• Xn 2Xn-1 - Xn-2

13
Entropy Coding

• Removes redundancy from residual signal
• Removes redundancy from the DCT coefficients
• No lost information
• Huffman, RLE, Rice

14
Bibliography

• Signal to Noise in Documentation, Michael J.
  Albers, University of Memphis, Tennessee, 2004
• M. Alexander Broadhead and Charles B. Owen,
  Dartmouth College, Hannover New Hampshire, 1995,
  International Multimedia Conference
• Sound Analysis Using MPEG Compressed Audio,
  George Tzanetakis and Perry Cook, Princeton
  University, Princeton New Jersey

15
Bibliography 2

• Perceptual Cryptography on MPEG Layer 3
  Bit-Streams, Andres Torrubia and Francisco Mora
• A New Perceptual Quality Measure for Bit Rate
  Reduced Audio, Thilo Thiede and Ernst Kabot,
  Technical University of Berlin, Germany
• Lossless Compression of Digital Audio, Mat Hans
  and Ronald W. Schafer