Ch. 5: Audio Compression Multimedia Systems

Transcription

1 Ch. 5: Audio Compression Multimedia Systems Prof. Ben Lee School of Electrical Engineering and Computer Science Oregon State University Chapter 5: Audio Compression 1 Introduction Need to code digital audio efficiently and with quality: Storage of high fidelity audio: CD, DAT, MD Interpersonal communication: video conferencing, multimedia applications... Audio standards for HDTV. Digital sound broadcasting. Two types of codecs: Waveform coding (time domain). Perceptual coding (frequency domain). Chapter 5: Audio Compression 2 1

2 Outline Sampling techniques (PCM): Linear Non-linear Generic Coding Techniques: DPCM ADPCM Psychoacoustic Coding: MPEG Chapter 5: Audio Compression 3 Audio Coding Standards Standard IMA- ADPCM G.711 G.722 G.728 G.723 Audio CD MPEG-1 BW (Hz) Compression ADPCM!-law or A- law PCM DPCM low-delay CELP ADPCM Linear PCM sub-band coding Sample Rate (KHz) Precision (bits) Bit rate (kbps) or (stereo) IMA: Interactive Multimedia Association CELP: Code excited linear prediction Quality Telephone, CD VoIP VoIP VoIP Video Conferencing CD near-cd Chapter 5: Audio Compression 4 2

3 Pulse Code Modulation Quantization error ( noise ) Each bit of resolution adds 6 db of dynamic range. 16-bit samples has 96 db SNR. Greater the number of quantization levels, lower the quantization noise Tradeoff between quantization noise performance vs. data rate. Number of bits required depends on the amount of noise that is tolerated 6.02N SNR Chapter 5: Audio Compression 5 Linear Quantization Quantization levels are evenly spaced. 16-bit samples provide plenty of dynamic range. CDs do this. Compression factor 1:1. File formats.wav (MS).AIFF (Unix and Mac) but SNR is worse for low-level signals than for high-level signals. Chapter 5: Audio Compression 6 3

4 Sound Examples 4-bit 8-bit 16-bit 16-bit voice 5 KHz 22 kbps 44 kbps 80 kbps 84 kbps 11 KHz 40 kbps 80 kbps 152 kbps 160 kbps 22 KHz 80 kbps 152 kbps 296 kbps 308 kbps 44 KHz 152 kbps 296 kbps 1.1 Mbps (stereo) 612 kbps Chapter 5: Audio Compression 7 Non-linear Quantization Humans are less sensitive to changes in loud sounds than quiet sounds Non-linear quantization of the signal s amplitude: Quantization step-size decreases logarithmically with signal level. Low-amplitude samples represented with greater accuracy than highamplitude samples. Logarithmic-compressed quantizer. Chapter 5: Audio Compression 8 4

5 Non-linear Sampling Output Input Chapter 5: Audio Compression 9!-Law ( mu-law ) Companding ln(1 +! x ) f(x) = 127 sign(x) ln(1 +!) (x normalized to [-1, 1]) Typically! = 255 Chapter 5: Audio Compression 10 5

6 !-Law Companding Provides 14-bit quality (dynamic range) with an 8-bit encoding Used in North American & Japanese ISDN voice service Simple to compute encoding. Compression factor 2:1. Example file formats:.wav (Microsoft),.au (Sun). Chapter 5: Audio Compression 11!-Law Encoding High-resolution PCM encoding (12, 14, 16 bits) x = bit discarded Table Lookup Sender! = bit!-Law encoding Codes the absolute value of the input (with bias of 33) in signed-magnitude. Valid samples => to 8158 Chord Range 33(0) 63(30) 64(31) 127(94) 128(95) 255(222) 256(223) 511(478) 512(479) 1023(990) 1024(991) 2047(2014) 2048(2015) 4095(4062) 4096(4061) 8191(8158) Sign bit removed Step Size Chord => finding the most significant 1 bit Step => the four bits following the most significant 1 bit Chapter 5: Audio Compression 12 6

7 !-Law Decoding 8-bit!-Law encoding Inverse Table Lookup Receiver 14-bit decoding Chapter 5: Audio Compression 13 Outline Sampling techniques (PCM) Linear Non-linear Generic Coding Techniques DPCM ADPCM Psychoacoustic Coding MPEG Chapter 5: Audio Compression 14 7

8 Differential PCM (DPCM) Exploit temporal redundancy in samples. Difference between two n-bit samples can be represented with significantly fewer than n-bits Transmit the difference (rather than individual samples). 2:1 ~ 4:1 compression ratio. Chapter 5: Audio Compression 15 DPCM Signal to be encoded Δx = x x ˆ Prediction Difference/Error x + Δx Δ x + Quantizer Entropy Coder x-bit DPCM difference x ˆ Predictor x Predicted Signal Reconstructed Signal x ˆ (n) = x (n 1) Simple difference x = x ˆ + Δ x N x ˆ (n) = h k x (n k) Linear predictor => minimizes quantization error k=1 x ˆ (n) = 0 PCM Chapter 5: Audio Compression 16 8

9 Slope Overload Problem Signal to be coded Dx(n) too large for quantizer to handle Predicted signal Occurs especially near the Nyquist frequency. Error introduced leads to severe distortion. Chapter 5: Audio Compression 17 Adaptive DPCM (ADPCM) Use a larger step-size to encode differences between high-frequency samples & a smaller step-size for differences between low-frequency samples. Use previous sample values to estimate changes in the signal in the near future. Chapter 5: Audio Compression 18 9

10 ADPCM To ensure differences are always small... Adaptively change the step-size (quanta). (Adaptively) attempt to predict next sample value. y-bit PCM sample x + + x ˆ Δx Difference Quantizer Step-Size Adjuster Δ x x-bit ADPCM difference Predictor x + + Δ x + Dequantizer x = x ˆ + Δ x Chapter 5: Audio Compression 19 Outline Sampling techniques (PCM) Linear Non-linear Generic Coding Techniques DPCM ADPCM Psychoacoustic Coding MPEG Chapter 5: Audio Compression 20 10

11 Psychoacoustic Principles Based on extensive studies of human perception. Average human does not hear all frequencies the same way. Limitations of the human sensory system leads to facts that can be used to cut out unnecessary data in an audio signal. Two main properties of the human auditory system: Absolute threshold of hearing Auditory masking Chapter 5: Audio Compression 21 Time vs. Frequency Interpretation H(W) time h(t) Chapter 5: Audio Compression 22 11

12 Absolute Threshold of Hearing Threshold in quite T q (f) T q ( f ) = 3.64( f /1000) e 0.6( f / ) ( f /100) 4 Range is about 20 Hz - 20 khz, most sensitive at 2 khz to 4 khz. Dynamic range (quietest to loudest is about 96 db) Normal voice range is about 500 Hz - 2 khz. Chapter 5: Audio Compression 23 Simultaneous Masking The presence of tones at certain frequencies makes us unable to perceive tones at other nearby frequencies Humans cannot distinguish between tones within 100 Hz at low frequencies and 4 khz at high frequencies. Chapter 5: Audio Compression 24 12

13 Masking Threshold Approximates a triangular function modeled by spreading function, SF(x) (db), where x has units of bark SF(x) = (x ) (x ) 2 Chapter 5: Audio Compression 25 Simultaneous Masking Example Listen to these played together Signal Noise Signal + Noise (SNR = 24 db) Chapter 5: Audio Compression 26 13

14 Critical Bands(1) Human auditory Limited and frequency dependant resolution 25 critical bands Bark: Barkhausen 1 Bark = width of one critical band Critical band number (Bark) for a given frequency, z(f): f < 500Hz => z(f) f/100 f > 500Hz => z(f) log 2 (f/1000) Also given as z( f ) = 13.0arctan( 0.76 f ) + 3.5arctan( f 2 /56.25) Chapter 5: Audio Compression 27 Critical Bands(2) Chapter 5: Audio Compression 28 14

15 Temporal Masking If we hear a loud sound, then it stops, it takes a little while until we can hear a soft tone nearby. As much as 50 ms before and 200 ms after. Example: Play 1 khz masking tone at 60 db and 1.1 khz test tone at 40 db. SPL (db) Pre-masking 60 simultaneous Post-masking t (ms) Chapter 5: Audio Compression 29 Global Masking Threshold Additive process The minimum level of audibility in the presence of masking noise Time varying function of frequency at each frequency at a given time Audio Level 500Hz Masker 200Hz Masker 2KHz Masker 5KHz Masker Masker Threshold KHz Maskee Frequency Chapter 5: Audio Compression 30 15

16 Summary Auditory masking Range of masking is 1 Bark (or critical band) Two factors for masking - frequency masking and temporal masking. How can we use this information for compression? Chapter 5: Audio Compression 31 MPEG-1 Audio Founded in 1988 (ISO/IEC/ JTC 1/SC29/WG11) Audio and video standard for encoding and decoding. Specify general functionality of application. MPEG-1: 1.5 Mbps for audio and video 1.2 Mbps for video 0.3 Mbps for audio Uncompressed CD audio => 44,100 samples/sec 16 bits/sample 2 ch > 1.4 Mbps Compression ratio from 2.7:1 to 42:1 16 bit stereo sampled at 48 khz is reduced to 256 kbps Sampling frequency 32, 44.1 and 48 khz One or two audio channels Monophonic, Dual-monophonic, Stereo, Joint Stereo Chapter 5: Audio Compression 32 16

17 MPEG-1 Audio Layers Layer 1 DCT type filter with one frame and equal frequency spread per band. Psychoacoustic model only uses frequency masking. Layer 2 Use three frames in filter (before, current, next, a total of 1152 samples). This models a little bit of the temporal masking. Layer 3 Better critical band filter is used (non-equal frequencies) Psychoacoustic model includes temporal masking effects, and takes into account stereo redundancy. Huffman coder. Known as MP3 Chapter 5: Audio Compression 33 MPEG Codec Block Diagram PCM Audio Samples 32 Subband Filter Bit allocation, Quantization & Coding Encoding of Bitstream Encoded Bitstream Spectral Analysis & Psychoacoustic Model Decoding of Bitstream Per Band Masking Margins (Signal-to-Mask Ratio) Inverse Quantization Synthesis Filterbank Encoded Bitstream Chapter 5: Audio Compression 34 17

18 Sub-band Filtering 32 PCM samples yields 32 subband samples. Each sub-band corresponds to a frequency band evenly spaced from 0 to Nyquist frequency. For khz sampling rate, each sub-band is 750 Hz wide. Samples out of each filter are grouped into blocks, called frames. Blocks of 12 for Layer 1 (384 samples). Blocks of 36 for Layers 2 and 3 (1152 khz, represents 8 ms of audio. Time Frequency Chapter 5: Audio Compression 35 Subbands vs. Critical Bands 32 constant-width subbands do not accurately reflect the ear s critical bands Chapter 5: Audio Compression 36 18

19 Psychoacoustic Model Psychoacoustic Model 1 Applied to MPEG Layer I and II Lower complexity Psychoacoustic Model II Applied to MPEG Layer III Higher complexity Chapter 5: Audio Compression 37 Psychoacoustic Model I Implementation FFT gets detailed spectral information about the signal: 512-point FFT for Layer point FFT for Layer 2 and 3 From each subband s samples, separate tonal (sinusodial) and nontonal (noise) maskers in the signal. Determine masking threshold for each sub-band. Determine the global masking threshold. Determine the minimal masking threshold in each subband. Calculate the signal-to-mask ratio (SMR) in each sub-band, which is the ratio of signal energy to minimum masking threshold. Chapter 5: Audio Compression 38 19

20 Step 1: Spectral Analysis - FFT Transforms PCM samples from time to frequency domain. Needs finer frequency resolution for an accurate calculation of the masking thresholds. Incoming audio sample s(n) normalized according to: x(n) = s(n) N2 b 1 N - FFT length b - number of bits per sample Segments the signal into 512-sample frames => 12 Chapter 5: Audio Compression 39 Fast Fourier Transform FFT efficiently computes DFT: N 1 x[n]e j 2πnk X[k] = N k = 0,...,N 1 n= 0 e jx = cos x + j sin x Euler s formula Basic idea => separate X[k] into even, Y[k], and oddnumbered, Z[k], samples. X[k] = Y[k]+ e j 2π k N Z[k] for k = 0,1,...,N /2 1 X[k + N /2] = Y[k] e j 2π k N Z[k] for k = 0,1,...,N /2 1 Chapter 5: Audio Compression 40 20

21 Data Windowing (1) DFT inherently assumes that data is a single period of a periodically repeating waveform! Sampled values of the signal are multiplied by a (window) function that tapers toward zero at either end. Allows the sampled signal, rather than starting and stopping abruptly, "fades" in and out. Reduces the effect of the discontinuities where the mismatched sections of the signal join up. Chapter 5: Audio Compression 41 Data Windowing (2) Hann Window w[n] = 1 * $ 2πn '-, 1 cos& )/ 2 + % N (. Chapter 5: Audio Compression 42 21

22 Data Windowing (3) w[n] = 1 * $ 2πn '-, 1 cos& )/ 2 + % N (. 1/16 overlapped window t 512 samples (12 ms frame) 448 samples (10.9 ms) Chapter 5: Audio Compression 43 Spectral Analysis: PSD Power Spectral Density, P(k), using 512-point FFT (N=512) P(k) = db +10 log Power Normalization Term N 1 w(n)x(n)e j 2πkn N n= 0 Hann Window PSD FFT Conversion to db 2 db 0 k N 2 Generates N/2 frequency components or bins. Chapter 5: Audio Compression 44 22

23 PSD Example CD-quality pop music khz, 16 bits per sample Tonal masker x Noise masker o Chapter 5: Audio Compression 45 Step 2: Identify Tonal & Noise Maskers (1) Tonal maskers are signals that generate pure tone, i.e., harmonically rich. Noise masker has no single dominant freq., i.e., more noise like. Tonal and noise maskers have different masking characteristics. Component considered tonal if P(k)-P(k k ) 7 db, where k is Bark distance. k varies as function of frequency range Noise Masker 4 db X Tone Masker 24 db SPL (db) SPL (db) Critical Band f Critical Band f Chapter 5: Audio Compression 46 23

24 Step 2: Identify Tonal & Noise Maskers (2) Tonal maskers, P TM (k) P TM (k) = 10log 10 j= 1 0.1P(k+ j) P(k) - Output of FFT Energies from 3 adjacent spectral components centered around the peak are combined. Noise maskers, P NM (k) P NM (k) = 10log 10 j 1 0.1P( j) Energies from all spectral components within each critical band are combined. Chapter 5: Audio Compression 47 Step 3: Decimation Invalid Maskers Only retain maskers that satisfy P TM, NM (k) T q (k) Sliding 0.5 Bark-window is used to replace any pair of maskers occurring within a distance 0.5 Bark by the stronger of the two. P TM, NM (k) =arg max &' )*.,,*., P TM, NM (k+k 0 ) Chapter 5: Audio Compression 48 24

25 Step 4: Calculation of Individual Masking Threshold (1) Piecewise linear function and approximates the spreading function Slope = SF(x) = (x ) (x ) Slope =-17 Spreading Function utilized in Psychoacoustic Model 1: % 17Δ z 0.4P TM,NM ( j)+11, - 3 Δ z < 1 '(0.4P TM,NM ( j)+ 6)Δ z, -1 Δ z < 0 SF(i, j) = & ' 17Δ z, 0 Δ z <1 ( '(0.15P TM,NM ( j) 17)Δ z 0.15P TM,NM ( j), 1 Δ z < 8 Δ z = z(i) z( j) Chapter 5: Audio Compression 49 Step 4: Calculation of Individual Masking Threshold (2) Tonal masking threshold, T TM (i, j) T TM (i, j) = P TM ( j) 0.275z( j)+ SF(i, j) P TM (j) - SPL of the tonal masker in frequency bin j z(j) - Bark frequency of frequency bin j SF(i, j) - spread of masking from bin j to maskee bin i Noise masking threshold, T NM (i, j) T NM (i, j) = P NM ( j) 0.175z( j)+ SF(i, j) P NM (j) - SPL of the tonal masker in frequency bin j z(j) - Bark frequency of frequency bin j SF(i, j) - spread of masking from bin j to maskee bin i Chapter 5: Audio Compression 50 25

26 Individual Masking Threshold Example Tonal maskers Noise maskers Chapter 5: Audio Compression 51 Step 5: Calculation of Global Masking Threshold Global masking threshold, T g (i) # T g (i) = 10log 10 (0.1T q (i)) L + 10 (0.1T TM (i,l)) M + 10 (0.1T NM (i,m)) & % ( $ ' l=1 where i represents frequency index each masking threshold L - number of tonal maskers M - number of noise maskers m=1 Chapter 5: Audio Compression 52 26

27 Global Masking Threshold Example P TM, NM (k) T q (k) x 0.5Bark Window Chapter 5: Audio Compression 53 Step 6: Calculate Minimal Masking Threshold Minimal Masking Threshold in subband n (MMT(n)) MMT(n)= min 0 1(3) $ %& ' )*++,-. / Chapter 5: Audio Compression 54 27

28 Calculate Signal to Mask Ratio Minimal Masking Threshold in subband n (MMT(n)) MMT(n)= min 1 2(4) % &' ( *+,,-./ 0 Chapter 5: Audio Compression 55 Spread of Masking A masker centered within one critical band has some predictable effect on detection thresholds in other critical bands. Chapter 5: Audio Compression 56 28

29 Masking Model Example Suppose the levels of the first 16 of the 32 sub-bands are: Band Level (db) The level of the 8th band is 60 db, the pre-computed masking model specifies a masking of 12 db in the 7th band and 15 db in the 9th. The signal level in 7th band is 10 (< 12 db), so ignore it. The signal level in 9th band is 35 (> 15 db), so send it. Only the signals above the masking level needs to be sent. Chapter 5: Audio Compression 57 Scaling Block of 12 samples for each subband is scaled to normalize the peak signal level within a subband: Largest signal quantized using 6-bit scale-factor. The receiver needs to know the scale factor and quantisation levels used: Information included along with the samples The resulting overhead is very small compared with the compression gains. Chapter 5: Audio Compression 58 29

30 Bit Allocation For each audio frame, bits must be distributed across the sub-bands from a predetermined number of bits defined by the target bit rate. 192 kbps target rate => 8 => 1.54 kbits/frame. Objective is to minimize noise-to-mask ratio (NMR) over all sub-bands, i.e., minimize quantization noise. Many different variations and optimizations. Chapter 5: Audio Compression 59 Quantization Noise Equivalent to x[n] = Q[x[n]] + e[n] SNR= 20log(x[n]/e[n]) More bits allocated to quantization, smaller e[n]! Chapter 5: Audio Compression 60 30

31 Bit Allocation SNR increases resolution by 6.02 db for each bit added SNR = 20 log V V QN = 20 log 2 N 1 1/2 = 6.02N(dB) SMR determined by Tonal or Noise masking threshold NMR increases for each bit added Iterative process: For each sub-band, determine NMR = SNR SMR Allocate bits one-at-a-time to the sub-band with the smallest NMR. Recalculate SNR, and thus NMR values. Iterate until all bits used. Chapter 5: Audio Compression 61 Bitstream Organization Encoded Samples 384 for Layer 1, 1152 for Layers 2 and 3 Side Info Encoding parameters (Layer specific) Frame Header (32 bits) Syncword Bit-rate and sampling information. Number of channels (1 or 2) Misc. other stuff. Chapter 5: Audio Compression 62 31

32 Frame Header Sampling rate frequency index khz khz khz 11 - reserved Pad bit 0 - frame is not padded 1 - frame is padded with one extra slot Private bit Channel Mode 00 - Stereo 01 - Joint stereo (Stereo) 10 - Dual channel (2 mono channels) 11 - Single channel (Mono) Frame Sync (set to 1s) Layer description 00 - reserved 01 - Layer III 10 - Layer II 11 - Layer I MPEG Audio version ID 00 - MPEG Version 2.5 (later extension of MPEG 2) 01 - reserved 10 - MPEG Version 2 (ISO/IEC ) 11 - MPEG Version 1 (ISO/IEC ) Bit rate index Layer 1: kbps Layer 2: kbps Layer 3: kbps Misc. Copyright Original Protection bit 0 - Protected by CRC (16-bit CRC follows header) 1 - Not protected Chapter 5: Audio Compression 63 Frame Structure (0xfff0) sync Framelen Bit-allocation Scale-factors Audio Data 16-bit 16-bit 4x32=128-bit 6xN=>192-bit (max.) 12 samples x N Framelen - frame length in bytes; Bit-allocation - 4 bits allocated to each subband (0-31) bits, 0000 indicates no sample. Scale-factors - 6 bits Multiplier that sizes the samples to fully use the range of the quantizer Has variable length depending on N = the number of subbands with non-zero bit allocation. Audio Data - subbands 0-31 stored in order, with each subband including 12 samples Chapter 5: Audio Compression 64 32

33 MPEG Audio Layer 3 (MP3) Layer samples per frame Up to 448 kbps Layer samples per frame Up to 384 kbps Layer 3 (MP3) Range of bit-rates from 8 kbps to 320 kbps Better filterbank Improved psychoacoustic model Better bit-allocation process Entropy coding... Chapter 5: Audio Compression 65 MP3 Block Diagram Chapter 5: Audio Compression 66 33

34 Improvements Equally spaced Sub-bands do not accurately reflect the ear s critical bands: f < 500Hz => z(f) f/100 f > 500Hz => z(f) log 2 (f/1000) Each sub-band further analyzed using Modified DCT (MDCT) to create 18 samples (for total of 576 samples). Increases the potential for redundancy removal. Better tracking of masking threshold. MP3 also specifies a MDCT block length of 6. Lots of bit allocation options for quantizing frequency coefficients. Quantized coefficients Huffman coded. Chapter 5: Audio Compression 67 Effectiveness of MPEG Audio Layer I II III Target bit rate Layer I II III 192 kbps 128 kbps 64 kbps Complexity Complexities Encoder Decoder > Effectiveness of MPEG audio Ratio 64 kbits 128 kbits 4: :1 2.1 to :1 3.6 to 3.8 5: perfect, 4:just noticeable, 3 : slightly annoying, 2: annoying, 1: very annoying 4+ Theoretical Min. Delay 19 ms 35 ms 59 ms Chapter 5: Audio Compression 68 34

35 Questions Chapter 5: Audio Compression 69 35