Digital Audio Compression - Multimedia Computing - Lecture Slides, Slides of Multimedia Applications

Download Digital Audio Compression - Multimedia Computing - Lecture Slides and more Slides Multimedia Applications in PDF only on Docsity!

Digital Audio Compression

Speech Compression

• Compression of voice data
  • We have previously mentioned several methods that are used to compress voice data - mu-law and A-law companding - ADPCM and delta modulation
  • These are examples of methods which work in the time domain (as opposed to the frequency domain) - Often they are not even considered compression methods

General Audio Compression

• If we want to compress general audio (not

just speech), different techniques are

needed

• In particular, music compression is a more general form of audio compression
• We make use of psychoacoustical

modeling

• Enable perceptual encoding based upon an analysis of the ear and brain perceive sound
• Perceptual encoding exploits audio elements that the human ear cannot hear well

Psychoacoustics

• If you have been listening to very loud music,

you may have trouble afterwards hearing soft

sounds (that normally you could hear)

• Temporal masking
• A loud sound at one frequency (a lead guitar)

may drown out a sound at another frequency

(the singer)

• Frequency masking

Equal-Loudness Relations

Threshold of Hearing

• The following image is a plot of the threshold

of human hearing for pure tones – at loudness

below the curve, we don’t hear a tone

Frequency masking

• We can determine how a pure tone at a

particular frequency affects our ability to hear

tones at nearby frequencies

• Then, if a signal can be decomposed into

frequencies, for those frequencies that are

only partially masked, only the audible part

will be used to set the quantization noise

thresholds

Critical Bands

• Human hearing range divides into critical

bands

• Human auditory system cannot resolve sounds better than within about one critical band when other sounds are present
• Critical bandwidth represents the ear’s resolving power for simultaneous tones
• At lower frequencies the bands are narrower than at higher frequencies
• The band is the section of the inner ear which responds to a particular frequency

Critical Bands

• Generally, the audio frequency range for

hearing (20 Hz – 20 kHz) can be

partitioned into about 24 critical bands

(25 are typically used for coding

applications

• The previous slide does not show several of the highest frequency critical bands
• The critical band at the highest audible frequency is over 4000 Hz wide
• The ear is not very discriminating within a critical band

Temporal Masking

• A loud tone causes the hearing receptors in

the inner ear to become saturated, and they

require time to recover

• This leads to the temporal masking effect
• After the loud tone we cannot immediately hear another tone – post-masking - The length of the masking depends on the duration of the masking tone
• A masking tone can also block sounds played just before – pre-masking (shorter time)

MPEG Audio Compression

• MPEG (Motion Picture Experts Group) is a

family of standards for compression of both

audio and video data

• MPEG-1 (1991) CD quality audio
• MPEG-2 (1994) Multi-channel surround sound
• MPEG-4 (1998) Also includes MIDI, speech, etc.
• MPEG-7 (2003) Not compression – searching
• MPEG-21 (2004) Not compression – digital rights management

MPEG Audio Compression

• MPEG-1 defined three downward

compatible layers of audio compression

• Each layer offers more complexity in the psychoacoustic model used and hence better compression
• Increased complexity leads to increased delay
• Compatibility achieved by shared file header information
• Layer 1 – used for Digital Audio Tape
• Layer 2 – proposed for digital audio broadcasting
• Layer 3 – music (MPEG-1 layer 3 == mp3)

MPEG Audio Compression

• PCM input filtered into 32 bands
• PCM FFT transformed for PA model
• Windows of samples (384, 576, 1152) coded

at a time

MPEG Audio Compression

• Since the sub-bands overlap, aliasing may

occur

• This is overcome by the use of a quadrature mirror filter bank - Attenuation slopes of adjacent bands are mirror images