Efficient Representation of Sound Images: Recent Developments in Parametric Coding of Spatial Audio

Transcription

1 Efficient Representation of Sound Images: Recent Developments in Parametric Coding of Spatial Audio Dr. Jürgen Herre 11/07 Page 1 Jürgen Herre für (IIS) Erlangen, Germany

2 Introduction: Sound Images? Humans live in a world of sound continuously listen to the sound surrounding us perceive the acoustic waves reaching the ears and interpret them (re)construct an acoustic scene In analogy to visual perception, we form a sound image... This talk: How to efficiently represent and reproduce sound images! Dr. Jürgen Herre 11/07 Page 2

3 Overview Part I Part II Part III What constitutes a sound image? Efficient coding of spatial sound images Perceptual audio coding Coding of multi-channel / surround sound Spatial Audio Coding & MPEG Surround Next-generation interactive coding / rendering of sound images: Spatial Audio Object Coding Demonstration Dr. Jürgen Herre 11/07 Page 3

4 Part I: What Constitutes A Sound Image? Dr. Jürgen Herre 11/07 Page 4

5 Spatial Sound Perception The Physics Part We are permanently listening to sound through our two sensors (ears) Ears receive a complex sound mixture of direct sound that is radiated from several sources (frequently concurrently), and many reflections from our acoustic environment ( ambient sound ) Dr. Jürgen Herre 11/07 Page 5

6 Spatial Sound Perception (2) The Perceptual Part Some Key Words Humans interpret sound to make sense of it Map it to an internal sound scene which is a perceptual correlate of the actual physical scene (not necessarily identical!) Very complex & sophisticated process - by far not fully understood! Psychoacoustics [Zwicker, Moore, Fletcher, Blauert, ] Perceptual streaming, auditory scene analysis [Bregman, ] How to form a consistent interpretation of the acoustic world from the stream of received information fragments Dr. Jürgen Herre 11/07 Page 6

7 Spatial Sound Perception (3) A Note of Caution In the following, we will look at the phenomenon of spatial audio perception in a grossly simplified, yet illustrative and very useful way Allows immediate technical application to efficient coding of spatial sound Highlight some analogies between sound images and visual images Frequency Selectivity The Human Auditory System processes sound in a frequency selective way (roughly logarithmic, ca. 1/3 octave BARK and ERB scales) Dr. Jürgen Herre 11/07 Page 7

8 Spatial Sound Perception (4) Sound Localization Complex process, mostly driven by interaural signal differences L R Depending on its position, a sound source produces differences in the ear signals Most important perceptual cues contained in the ear signals: e.g. [Blauert 1984] Interaural Level Differences (ILD) Interaural Phase/Time Differences (IPD/ITD) (relevant up to ca. 4 khz frequency) Temporal signal envelope (high frequencies) ILD, IPD/ITD determine perceived lateral position of sound sources! Dr. Jürgen Herre 11/07 Page 8

9 Spatial Sound Perception (5) In realistic closed rooms, reflections from the walls (and other objects) occur These blur the simple ILD, IPD/ITD relations by decorrelating the ear signals Decorrelated sound at the two ears represents reflections and thus provides information on the acoustic environment Distinct early reflections nearby walls Late reverberation statistical summary of acoustic environment / room properties Auditory system needs to parse sound into direct & ambient sound! Dr. Jürgen Herre 11/07 Page 9

10 Spatial Sound Perception (6) e1 [Faller 2004] Dr. Jürgen Herre 11/07 Page 10 e2 Important role of Interaural Coherence (IC): Determines spatial extent (=width) of sound event [Schuijers03] [Faller03] Perceived width of auditory event increases as coherence decreases (1 4); eventually separates into 2 distinct events Max. of normalized cross-correlation: IC = max d $ % n=#$ $ e 1 (n) " e 2 (n + d) % e 2 1 (n) " % e 2 2 (n + d) Key to source width and source/ambience perception! n=#$ $ n=#$

11 Some A V Analogies in Scene Attributes Visual Domain Auditory Domain ( auditory cues) Foreground object Sound sources ( high IC) - directional Object position - Object position ( ILD/ITD/IPD for lat. position) Object size - Object size ( IC) Background Ambience ( low IC) - non-directional Dr. Jürgen Herre 11/07 Page 11

12 Part II: Efficient Coding Of Spatial Sound Images Dr. Jürgen Herre 11/07 Page 12

13 Basics of Audio Coding Goal Predominant Approach Represent audio data as compactly as possible while maintaining sound quality (ideally: transparent coding) Concept of perceptual audio coding Optimize subjective quality rather than objective distortion metrics (e.g. MSE/SNR) Use knowledge about signal receiver: Psychoacoustics gives limits of perception Keep coding distortion below limits! No universal source model available (unlike in speech coding) Dr. Jürgen Herre 11/07 Page 13

14 Psychoacoustics 80 db L T Masking Threshold Masker 20 Threshold in quiet Inaudible Signal Masked Sound 0 0,02 0,05 0,1 0,2 0, khz f T Dr. Jürgen Herre 11/07 Page 14

15 Demonstration: The "13 db Miracle" Historic demonstration by James D. Johnston and Karlheinz Brandenburg at AT&T Bell Laboratories in 1990 using the best psychoacoustic model available at that time... Original signal Original + white noise, SNR = 13,6 db Original + noise at threshold, SNR = 13,6 db Difference signal: White noise Difference signal: Noise at threshold SNR does not adequately describe subjective sound quality! Putting psychoacoustics to work makes a huge difference! Dr. Jürgen Herre 11/07 Page 15

16 Basic Paradigm of (Monophonic) Perceptual Audio Coding audio in Analysis Filterbank Quantization & Coding Encoding of Bitstream bitstream out Perceptual Model bitstream in Decoding of Bitstream Inverse Quantization Synthesis Filterbank audio out Dr. Jürgen Herre 11/07 Page 16

17 A Real Audio Coder (MPEG-2 AAC, 1997) Bitstream Output Bitstream Multiplexer Input time signal Gain Control Perceptual Model Filter Bank TNS Intensity/ Coupling Prediction M/S Scale Factors Quant. Rate/Distortion Control Noiseless Coding Dr. Jürgen Herre 11/07 Page 17

18 The Better Spatial Sound Image: Surround Sound / Multi-Channel Audio Significantly increased spatial realism over stereo, envelopment Origins in movie sound (5.1); now also for music, broadcasting Increasingly adopted in consumers homes Dr. Jürgen Herre 11/07 Page 18 Fraunhofer Institut

19 Traditional Delivery Formats For Surround Sound Matrixed Surround (Prologic, Neo6, ) Discrete Surround (AAC, AC-3, ) Downmix of 5.1 sound into stereo signal, upmix at the receiver side Efficient in terms of transmission bandwidth (same bitrate as stereo) Backward compatible to stereo delivery Limited computation necessary Significant loss in subjective audio quality Separate transmission of each channel Significantly higher bitrate than stereo Moderate amount of computation High subjective audio quality possible Dr. Jürgen Herre 11/07 Page 19

20 A Major Step Ahead: Spatial Audio Coding Rather recent development Main Characteristics Compression efficiency: Transmits multi-channel audio at bitrates used for 2-channel stereo (or even mono) Backward compatibility: SAC multi-channel audio is coded in a backward compatible way existing infrastructures can be seamlessly upgraded to multi-channel / surround! High subjective audio quality Heavily based on exploiting perception rather than waveform coding! Dr. Jürgen Herre 11/07 Page 20

21 The Spatial Audio Coding (SAC) Concept Spatial Audio Coding = Downmix + Parametric Spatial Synthesis e.g. ICLD, ICTD/ICPD, ICC Extracts & reproduces Inter-Channel Counterparts of Inter-Aural Parameters Dr. Jürgen Herre 11/07 Page 21

22 Related Technology Generalization / Extension of Binaural Cue Coding Parametric Coding of Multi-Channel into a Mono Channel Parametric Stereo Parametric Coding of 2-Channel Stereo into a Mono Channel Matrix Surround (Dolby Prologic, Logic 7, Circle Surround ) First commercial Application (2003) MP3 Surround Backward compatible extension of stereo MP3 towards 5.1 surround sound at kbit/s Dr. Jürgen Herre 11/07 Page 22

23 Example: MP3 Surround Encoding Principle: Surround Stereo + Side Information L R C Ls Rs Compatible Downmix Lc Rc Standard MP3 Encoder MP3 Surround Bitstream Surround BCC Encoder Ancillary Data Surround Enhancement [Herre et al. 2004] Dr. Jürgen Herre 11/07 Page 23

24 Example: MP3 Surround Decoding Principle: Stereo + Side Information Surround MP3 Surround Bitstream Standard MP3 Decoder Lc Rc Advanced Surround Stereo Multi- Playback Channel (simple Decoderdecoder) Decoding L R C Ls Rs Ancillary Data Surround Enhancement [Herre et al. 2004] Dr. Jürgen Herre 11/07 Page 24

25 MPEG Spatial Audio Coding / MPEG Surround Work item Spatial Audio Coding (SAC) Technical development 3/2004-7/2006 Main contributors: Fraunhofer IIS, Agere Systems, Coding Technologies and Philips Renamed into MPEG Surround (MPEG-D) Applications: Efficient & backward compatible upgrade of audio distribution to multi-channel, e.g.: Music download service Multi-channel streaming / Internet radio Digital audio broadcasting Dr. Jürgen Herre 11/07 Page 25

26 MPEG Surround Synthesis: General Concept Decoder Structure Input 1 hybrid analysis hybri d synthesi s Output 1 Input 2 hybrid analysis Spatial synthesis hybri d synthesi s Output 2 hybri d synthesi s Output N Spatial parameters hybrid = QMF filterbank + 2nd stage non-uniform frequency resolution relating to frequ. resolution of human auditory system Same QMF filterbank as in MPEG-4 HE-AAC (AAC + SBR) Dr. Jürgen Herre 11/07 Page 26

27 MPEG Surround Synthesis: General Concept (2) Frequency selective processing: Dynamic matrix + decorrelation Many more details, please see dedicated publications! Side information rate typ kbit/s Dr. Jürgen Herre 11/07 Page 27

28 Underlying Idea: Hierarchical En/decoding Encoder Coding Decoder l f l b OTT encoder element OTT decoder element l f l b c lfe r f r b OTT encoder element OTT encoder element TTT encoder element l 0 r 0 l 0 r 0 TTT decoder element OTT decoder element OTT decoder element c lfe r f r b Spatial parameters Spatial parameters Dr. Jürgen Herre 11/07 Page 28

29 MPEG Surround Concepts Generalization Similar trees for mono-based operation ( ), other modes ( 7-2-7, ) Spatial Parameters Channel Level Differences (CLDs) Inter-Channel Correlations (ICCs) Channel Prediction Coefficients (CPCs) Prediction errors (residuals) Other Aspects Decorrelation by QMF-domain all-pass filters Several tools for handling fine temporal envelope structure (both without and with additional side information) Dr. Jürgen Herre 11/07 Page 29

30 MPEG Surround: Additional Functionalities Artistic Downmix Matrix Surround Compatibility Enhanced Matrix Mode Binaural Rendering for headphone Externally created downmixes can be used Stereo downmix can be made compatible with common matrix surround decoders MPEG Surround decoder can decode matrix surround signal (i.e. work without side info) Downmix can be generated as virtual surround, or MPEG Surround can be decoded directly into virtual surround very efficiently Rich set of attractive features for practical application Dr. Jürgen Herre 11/07 Page 30

31 MPEG Surround: Recent Verification Test MPEG Surround MPEG Surround without side-info Matrixed Surround (Dolby Prologic II) Dr. Jürgen Herre 11/07 Page 31 Music-Store test scenario: Stereo downmix coded using

32 Part III: Next Generation Interactive Coding / Rendering of Sound Images From Spatial Audio Coding to Spatial Audio Object Coding Dr. Jürgen Herre 11/07 Page 32

33 Classic MPEG-4 Interactive Scene Composition (1996ff) audiovisual objects hierarchically multiplexed downstream control / data hierarchically multiplexed upstream control / data video compositor projection plane voice sprite 2D background scene coordinate system z y x audiovisual presentation user events 3D objects audio compositor voice Hierarchy of objects person sprite 2D background scene globe furniture desk audiovisual presentation Dr. Jürgen Herre 11/07 Page 33 hypothetical viewer speaker display user input Scene is composed of multiple A/V objects and can be rendered interactively

34 MPEG-4 Object Based (De)Coding Discrete approach comes at a rather high price: Bitrate and decoding complexity grow with number of objects Structural complexity Dr. Jürgen Herre 11/07 Page 34

35 From Spatial Audio Coding (SAC) to SAOC Regular Spatial Audio Coding: Channel-oriented scheme (MPEG Surround) Chan. #1 Chan. #2 Chan. #3 SAC Encoder Downmix signal(s) SAC Decoder Chan. #1 Chan. #2 Chan. #3 Chan. #4... Side Info Chan. #4... Dr. Jürgen Herre 11/07 Page 35

36 From SAC to SAOC (2) Alternative: Object-oriented Spatial Audio Coding Obj. #1 Obj. #2 Obj. #3 Obj. #4... SAOC Encoder Downmix signal(s) Side Info SAOC Decoder obj. #1 obj. #2 obj. #3 obj. #4... Renderer Chan. #1 Chan. #2... User Interaction / Ctrl. Processes object signals instead of channel signals Mixing /rendering parameters vary according to user interaction Combined obj. decoding & rendering computationally efficient! Previous work by Faller & Baumgarte [2001ff] and Faller [2006] Dr. Jürgen Herre 11/07 Page 36

37 Real-time interactive rendering of audio objects from a mono audio downmix + SAOC side information Dr. Jürgen Herre 11/07 Page 37

38 New MPEG Standardization Activities Work on Spatial Audio Object Coding (SAOC) started Transcoding approach: SAOC + rendering info MPEG Surround SAOC Decoder Reference model and working draft recently established (10/2007) Dr. Jürgen Herre 11/07 Page 38

39 Conclusions Sound Images carry some analogy to images in the visual world Spatial Audio Coding schemes code surround sound based on perception (rather than on waveform match) Object positions are represented by perceptual spatial parameters Audio Object Texture is coded using regular mono/stereo coder Such schemes can bring surround sound into existing infrastructures High compression factor (surround sound at 64kbps and below!) Stereo / mono backward compatibility Extension towards efficient interactive, object-based scene coding / rendering (Spatial Audio Object Coding) is currently on its way Dr. Jürgen Herre 11/07 Page 39

40 Dr. Jürgen Herre 11/07 Page 40