Thomas Schmidtschmidt@informatik.

haw-hamburg.de

Introduction to Audio Coding

o Fundamentals of Soundo Sampling & Quantizationo Compressiono Coding Standards

Thomas Schmidtschmidt@informatik.

haw-hamburg.deThe Nature of Sound

o Conversion of energy into vibrations in the air (or some other elastic medium)

o Most sound sources vibrate in complex ways leading to sounds with components at several different frequencies

o Sound is characterized at a very small timescale

o Frequency spectrum – relative amplitudes of the frequency components

o Range of human hearing: roughly 20Hz–20kHz, falling off with age

Thomas Schmidtschmidt@informatik.

haw-hamburg.dePsychoacoustics

Thomas Schmidtschmidt@informatik.

haw-hamburg.deWaveforms

o Sounds change over time- e.g. musical note has attack and decay, speech changes constantly

o Frequency spectrum alters as sound changes

o Waveform is a plot of amplitude against time

- Provides a graphical view of characteristics of a changing sound

- Can identify syllables of speech, rhythm of music, quiet and loud passages, etc

Thomas Schmidtschmidt@informatik.

haw-hamburg.deDigitization – Sampling

o Audio timescale implies minimum rate of 40kHz to reproduce sound up to limit of hearing- CD: 44.1kHz

- Sub-multiples often used for low bandwidth

- e.g. 22.05kHz for Internet audio

- DAT: 48kHz

- (Hence mixing sounds from CD and DAT will require some resampling, best avoided)

Thomas Schmidtschmidt@informatik.

haw-hamburg.deDigitization – Quantization

o 16 bits, 65536 quantization levels, CD quality

o 8 bits: audible quantization noise, can only use if some distortion is acceptable, e.g. voice communication

o Dithering – introduce small amount of random noise before sampling

- Noise causes samples to alternate rapidly between quantization levels, effectively smoothing sharp transitions

Thomas Schmidtschmidt@informatik.

haw-hamburg.deUndersampling & Dithering

Thomas Schmidtschmidt@informatik.

haw-hamburg.deData Size

o Sampling rate r is the number of samples per second, Sample size s bits

o Each second of digitized audio requires rs/8 bytes

o CD quality: r = 44100, s = 16, hence each second requires just over 86 kbytes (k=1024), each minute roughly 5Mbytes (mono)

o CD quality, stereo, 3-minute song requires over 25 Mbytes

Thomas Schmidtschmidt@informatik.

haw-hamburg.deCompression

o In general, lossy methods required because of complex and unpredictable nature of audio data

o Difference in perceiving sound and image means different approach from image compression:

+ Sampling & Quantization: PCM – DPCM – ADPCM

+ Special voice coding: Vocoders

+ Transform coding: DFT, DCT, MDCT

+ Perceptual coding

+ Entropy coding

Thomas Schmidtschmidt@informatik.

haw-hamburg.deCompanding

o Non-linear quantization

o Higher quantization levels spaced further apart than lower ones

o Quiet sounds represented in greater detail than loud ones

o µ-law, A-law (16 → 8 bit)

Thomas Schmidtschmidt@informatik.

haw-hamburg.deImproved Pulse Code Modulation

o Differential Pulse Code Modulation- Similar to video inter-frame compression

- Compute a predicted value for next sample, store the difference between prediction and actual value

o Adaptive Differential Pulse Code Modulation- Dynamically vary step size used to store quantized differences

Thomas Schmidtschmidt@informatik.

haw-hamburg.deSpeech Coders

There is a variety of well known ITU voice coders:

o G.711 – ‘the’ ISDN codec, µ / A-law companding, 8 bit@8kHz, (2:1)

o G.721, G.722, G.723

o G.726, G.727 – ADPCM with linear prediction, variable codewords 2 – 5 bit@8kHz (≤ 8:1)

o GSM 06.10 - ADPCM with linear prediction, 1,6bit@8kHz (10:1)

o Current compression values up to 50 : 1, 10 : 1 with decent quality (see www.speex.org)

Thomas Schmidtschmidt@informatik.

haw-hamburg.dePerceptually-Based Compression

o Identify and discard data that doesn't affect the perception of the signal- Needs a psycho-acoustical model, since ear and brain do not respond to sound waves in a simple way

o Threshold of hearing – sounds too quiet to hear

o Masking – sound obscured by some other sound

Thomas Schmidtschmidt@informatik.

haw-hamburg.deCompression by Masking

o Split signal into bands of frequencies using filters- Commonly use 32 bands

o Compute masking level for each band, based on its average value and a psycho-acoustical model

- i.e. approximate masking curve by a single value for each band

o Discard signal if it is below masking level

o Otherwise quantize using the minimum number of bits that will mask quantization noise

o Also: Temporal Masking

Thomas Schmidtschmidt@informatik.

haw-hamburg.deMPEG Audio Codingo MPEG Audio, Layer 3

o Three layers of audio compression in MPEG-1 (MPEG-2 essentially identical)

o Layer 1...Layer 3, encoding proces increases in complexity, data rate for same quality decreases

- Layer 2: temporal & frequency masking – Layer 3: MDCT + …

- e.g. Same quality 192kbps at Layer 1, 128kbps at Layer 2, 64kbps at Layer 3

o 10:1 compression ratio at high quality

o Variable bit rate coding (VBR)

Thomas Schmidtschmidt@informatik.

haw-hamburg.deReferences

• Z. Li, M. Drew: Fundamentals of Multimedia, Pearson Prentice Hall, 2004.• K. Rao, Z. Bojkovic, D. Milavanovic, Multimedia Communication Systems,

Prentice Hall, 2002. • N. Chapman, J. Chapman: Digital Multimedia, 2nd edition, Wiley,

Chichester, GB, 2004.