Issue link: http://digital.copcomm.com/i/1351516
16 S P R I N G 2 0 2 1 I C A S Q U A R T E R L Y Gordon Reid, Managing Director at CEDAR Audio. EMPHASIS In 1965, Ray Dolby invented the Dolby Noise Reduction System. His system is a form of signal processing designed to reduce tape hiss. Dolby Noise Reduction went through several iterations, Dolby A, B, C, etc. … but the foundation of the process is dynamic pre-emphasis. Considered an "encode-decode" system, program material is analyzed for volume and frequency content. Low-level passages and higher frequency audio, where noise is more present, are recorded at a higher volume to tape during the encode process to improve the signal to noise ratio. On the decode end, the volume and frequencies are reproduced at their original quieter level, and the noise floor gets quieter as well. The end result is not removing noise from the original signal, but rather a process to minimize additional noise from being introduced by the medium during playback. RESTORATION FOR RE-RELEASE By the late '80s and early '90s, the compact disc format was taking off and record companies knew there was a lot of money to be made by re-releasing their existing sound archives and libraries. However, there was a concern that many of these original recordings—many riddled with clicks, pops, and other signal degradation—might not be pristine enough to play alongside the new modern digital recordings of the day. In 1989, Gordon Reid was asked by the British Library National Sound Archive (BLNSA) to take research that it had funded within the Cambridge University Engineering Department, and help turn it into a commercial enterprise. Coming from an astrophysics background, he already had experience removing interference from radio telescope data. As it turned out, removing clicks and pops from a digital recording of an old record was not that different. After attending conference after conference and taking meetings all over the world, by the end of 1990, the first CEDAR systems were installed in major countries across the globe. Reid explains, "Sometimes described as 'removing the cream from the coffee,' audio restoration starts with a single signal that contains both the wanted signal and the unwanted noise. In the 1980's, the key to the problem was to design a set of filters that could remove the cream without ruining the coffee and make them efficient enough to run on the technology of the day. Different problems required different filters. De-clicking required signal modeling and interpolation. De-crackling required a new technology called 'Split & Recombine' that was invented in-house at CEDAR. And early de-hissing products used a technique called Spectral Subtraction." Other new filters tackled digital de-buzzing without creating the hollow sound of traditional filtering. CEDAR also introduced digital timing correction by inventing the dynamic sub-sample phase corrector. In other words, there is a lot of complex math going on under the hood! At the start, CEDAR and the other company active in the field at that time, Sonic Solutions, manufactured de-hissers based upon Spectral Subtraction that required a "noise print." A few years later, this idea was adopted by other companies such as Digidesign, which released the Digidesign Intelligent Noise Reduction (DINR) plugin for Pro Tools. From an advertisement I found online, "DINR reduces broadband noise, such as tape hiss, by 'intelligently' subtracting the noise from the digital audio file. First, the noise signature is created by selecting and analyzing an example of the noise within the source material. (A brief, isolated passage of noise is ideal but not necessary.) Using this signature, the effects of the noise are then removed or greatly reduced." This method of utilizing a noise print to reduce noise is still common in many plugins today. Noise print is still used in two modules of the CEDAR Cambridge System but, as Reid explains, that is more for "historical reasons." He continues, "Modern technology does not require a noise fingerprint. That's been superseded by various methods of estimating the noise at every given moment… So as the noise changes … it reevaluates what the noise content is." Introduced by CEDAR in an early form in 1994, but coming to prominence in its dialogue noise suppressors a few years later, this was perhaps our first sight of machine learning in audio processing. "It's like a dynamically updating fingerprint, even in the presence of the wanted signal," Reid explains. Since noise is, by definition, always changing, it makes sense that as the actual noise strays away from a traditional, static noise print, small bursts of noise get through or small bits of wanted signal get reduced, and the end result are the lovely artifacts that make us all cringe.