18/01/2024
Processing and noise reduction
Something I've just seen announced is a consumer level radio mic system with built in noise reduction processing. I expect this will prove very popular and more will follow.
These algorithms have been around for a while and have been using machine learning (or "AI") long before it was cool. They'll analyse the audio based on what the algorithm's been trained on as 'wanted audio' and 'unwanted audio'. Some algorithms remove unwanted frequencies with filters, which will remove them from the wanted signal too. Some reconstruct the audio based on what the algorithm expects it to be, which can result in errors appearing as additional frequencies or 'digital artefacts'
They've been around longest in post production and therefore they have the best tools- the largest selection, with the most control. Some are non-realtime (you can even draw on material to get rid of!), and if something doesn't work they get to have another go. They also have a controlled listening environment and will be able to hear things you can't wearing headphones in the field.
For a while now, some of these processing tools have been making their way out into the field, both as standalone hardware and have more recently been incorporated directly into some of our mixer/recorders as a paid software add-on.
Most (competent) production sound people will always supply an unprocessed signal. If this doesn't happen, the on set noise reduction gets 'baked in' and post production don't get to have another go, so the signal may be left with artefacts which weren't heard over headphones in a noisy environment. It's very easy to go 'too hard' on settings without realising.
In some situations where there is no post production, such as live broadcast- it's fine to have a 'baked in' signal, it may be more intelligible. However, it's usually incorporated into broadcast signal chains and doesn't need to be added closer to the source.
If you've got the power to use noise reduction, it might initially sound better in headphones, but post will have more tools at their disposal and they're going to get better results.
[Picture- Smiley face attenuated into a spectrogram on iZotope RX]
