Digital Audio RestorationDigital Audio Restoration

Simon GodsillSimon Godsill

Signal Processing GroupSignal Processing Group

University of CambridgeUniversity of Cambridge

www-sigproc.eng.cam.ac.uk/www-sigproc.eng.cam.ac.uk/~sjg~sjg

OverviewOverview

Audio Restoration - motivationAudio Restoration - motivation Audio Restoration in Cambridge 1984-2003Audio Restoration in Cambridge 1984-2003 Review of core technologiesReview of core technologies Audio restoration - principlesAudio restoration - principles Advanced topicsAdvanced topics Emerging techniquesEmerging techniques

Audio Restoration - Audio Restoration - motivationmotivation

Requirement to enhance material Requirement to enhance material fromfrom

– Sound Archives: Sound Archives: – Historical disk remastering:Historical disk remastering:– `Recent’ Magnetic Tape recordings: `Recent’ Magnetic Tape recordings:

– Forensic recordings, … Forensic recordings, …

Audio Restoration in CambridgeAudio Restoration in Cambridge

1984-88 - British Library funds1984-88 - British Library funds research into restoration of research into restoration of archived gramophone recordings at Signal Processing Group archived gramophone recordings at Signal Processing Group with Prof. Peter Rayner.with Prof. Peter Rayner.

1988 Cambridge company spun-out: CEDAR Audio. First real-1988 Cambridge company spun-out: CEDAR Audio. First real-time de-hiss and de-click in 1990, using DSP hardware on a PC time de-hiss and de-click in 1990, using DSP hardware on a PC platform.platform.

1990 -- Research into advanced audio processing at 1990 -- Research into advanced audio processing at Cambridge University - Godsill, Rayner, Wolfe, Fong, …Cambridge University - Godsill, Rayner, Wolfe, Fong, …

Core TechnologiesCore Technologies

De-click, de-crackleDe-click, de-crackle De-hissDe-hiss Resonant noise pulse removal Resonant noise pulse removal

De-click/de-crackle

De-click/de-crackleDe-click/de-crackle

Time domain models for clicks and Time domain models for clicks and audioaudio

Optimal detection and estimation Optimal detection and estimation of corrupted samplesof corrupted samples

Use fully Bayesian methods where Use fully Bayesian methods where time permitstime permits

De-hissDe-hiss

Frequency-domain methods Frequency-domain methods predominatepredominate

Non-linear processing of spectral Non-linear processing of spectral information to incorporate local information to incorporate local temporal and frequency dependencetemporal and frequency dependence

Time-domain model-based methods Time-domain model-based methods also developed (joint click/hiss also developed (joint click/hiss removal)removal)

*

* Courtesy Patrick Wolfe – see www-sigproc.eng.cam.ac.uk/~pjw47

*

Resonant noise pulsesResonant noise pulses

Tone-arm resonance in the presence Tone-arm resonance in the presence of breakages or other severe of breakages or other severe damage to gramophone disk groovesdamage to gramophone disk grooves

Simplest methods subtract an Simplest methods subtract an averaged template for the transientaveraged template for the transient

More sophisticated methods apply a More sophisticated methods apply a stochastic model for the resonant stochastic model for the resonant systemsystem

Low frequency noise pulse removal, contd.

Advanced TopicsAdvanced Topics

Bayesian statistical modelsBayesian statistical models De-clipping/de-quantizingDe-clipping/de-quantizing Pitch variation defects (Wow)Pitch variation defects (Wow)

Future DirectionsFuture Directions

ResourcesResources

`Digital Audio Restoration - a `Digital Audio Restoration - a statistical model-based approach’statistical model-based approach’ by Simon Godsill and Peter Rayner, by Simon Godsill and Peter Rayner, Springer-Verlag 1998Springer-Verlag 1998

See See www-sigproc.eng.cam.ac.uk/~sjg www-sigproc.eng.cam.ac.uk/~sjg for extracts, publications and for extracts, publications and sound examplessound examples