Digital Audio

1
Digital Audio

• Introducing perceptual encoders
• Using psychoacoustics

2
Principal features of the ear / brain response

• Ear is approximately logarithmic in subjective
  response to increasing volume (the 3dB fader in
  audio control)
• Response of the ear to a fixed audio spectral
  distribution (e.g an audio recording) is
  subjectively different as the volume changes (the
  phon curves) bass becomes more pronounced as
  volume increases (loudness controls) and
  richness of tone increases over about 60db
  loudness level (i.e. harmonics introduced by
  hearing system)

3
Principal features of the ear / brain response
part 2

• Masking occurs in both time and frequency
• in frequency, this modifies the effective
  threshold of hearing
• in time this can modify the effective threshold
  of hearing before a loud signal arrives (cuts
  off the build up) and after the signal stopped.
  The before deaf period can be a few msec, the
  after deaf period can be upto 200msec
• Equivalently, the ear behaves differently for
  long and short duration bursts i.e when
  compared to a few hundred msec.

4
Masking holds to key to psychoacoustic
codes..raise significantly, the effective
threshold of hearing
MAF- minimum audible field threshold of hearing
5
Even with simple MAF curves

• Audible dynamic range is less at 50Hz than
  5kHz..
• Maybe we could start then by splitting the audio
  spectrum into bands and quantising each one
  differently forget masking at this stage..

6
Sub band coding digital audio sampled at
48ksamples/sec
16x48000 768 kbps
4x3x16000 192 kbps (reduce bits Per sample)
16x3x48000 2304 kbps
Bit rates
16x3x16000 768 kbps (1/3 in each channel)
7
Reducing the bits per sample an example using
only the standard MAF
8
Bits needed in different parts of the spectrum
still no masking
80
Peak Signal Level
70
9 bits
9 bits
10 bits
10 bits
10 bits
9 bits
10 bits
11 bits
12 bits
11 bits
12 bits
12 bits
60
50
40
Sound Pressure Level dB-SPL
30
Threshold of Hearing
20
10
Frequency Hz
0
5000
10000
15000
-10
-20
-30
9
Using masking a psychoacoustic model..
Signal
Signal Noise (SNR 24 dB)
Noise

• Signal suppresses the noise
• Raises effective threshold of hearing to the
  masking threshold
• Establish a model of the new threshold of hearing
  and use this to determine the resolution needed
  in a particular frequency band essence of MPEG
  audio codes (but at present ignore temporal
  masking)
• Only use the dynamic range needed and make this
  adaptive

10
Single tone masking
11
Masking on a more complex signal
12
Other information needed in addition to the coded
audioframes

• The code is prepared and processed in frames of a
  predetermined length
• Each frame contains mostly coded audio, but in
  addition
• The peak level in each frequency sub band
• The masking level in each sub band
• The number of bits in each sample in each sub
  band

13
What does the system look like??- encoding
compressed audio
Digital Audio In
Sub-band filter bank
Scale and Quantise
Multiplex and Data Format
Coded Audio Out
Masking thresholds
Code additional Info
Psycho-acoustic model
FFT
ENCODER
Everything is done in the digital domain the
analogue original has been digitised to high
quality before entering the coder
14
What does the system look like??- decoding
compressed audio
15
MPEG 1 digital audio standards

• Three perceptual coders in the MPEG 1
  specification
• Layers 1, 2 3
• Layer 1 (.mp1)
• Similar to the simple coder just described
• 32 sub-bands are used
• Each frame contains 384 samples (32 x 12) lasting
  about 8msec
• A version of layer 1 was used in the Digital
  Compact Cassette (DCC)
• Layer 2 (.mp2)
• Slightly more complex but better quality than
  layer 1
• Frame length increased to 1152 samples (32 x 36)
  lasting about 20msec

16
MPEG 1 digital audio standards - more

• Layer 2 (continued)
• Data formatting of samples and side information
  is slightly more efficient
• Used in Digital Audio Broadcasting (DAB)
• Layer 3 (.mp3)
• Significantly more complex than layers 1 or 2
• Capable of reasonable quality even at very low
  data rates
• A combination of fixed sub-band coding and
  adaptive frequency transform coding is used to
  give up to 576 frequency bands (compared to 32
  for layers 1 2)
• Uses signal statistics as well as signal waveform
  for coding
• Huffman encoding is applied to samples (more on
  these to come..)
• MP3 files most prevalent of compressed
  audio.mp3 players etc
• Introduced late 1990s.

17
From wax discs to MP3.

• Original analogue records (discs, audio tape)
  attempted to cope with all possible signals at
  all possible times
• The CD (1980) did the same and included error
  correction (almost 1GByte/hour)
• Perceptual encoders eliminate what isnt relevant
  to the listener and compress to 100MByte per hour
  (and less) using MP3 and related formats