32-bit Float Audio - Clarity

I’ve been listening to the Sound Matters podcast, which I love, and have huge respect for both Simon and Chris and their glittering careers. They are truly excellent professionals who know their respective stuff, and the podcast offers some excellent perspectives on workflow and experiences at the very top tier of film work.

But I was lead to looking for some clarity on one thing that was part of the discussion on 32-bit float audio recording in one episode. It was said, I think by Chris Munro, that 32-bit float is like having more data to define each decibel of audio level. I think one of the analogies was that it was like having a fader with finer steps. I’m not sure if it was maybe an incorrect analogy or a poor selection of wording.

Either way I was surprised to hear this, as my understanding was that this was not the case, and was lead to do do some further reading around the topic. As I read, it became clear that the truth was a little more nuanced and complicated than I had previously understood. 

At this point I want to declare that, despite my courses occasional shortcomings, I am regularly thankful for my degree in Broadcasting Technology, in which we delved deep into the conversion of analog to digital audio (and video). Even though it was some 20 years ago now, that knowledge has always been an incredible base to learn new things from across this whole industry.

Let's start with what 32-bit float recording does

32-bit float records a file with dynamics well beyond the limits of the available microphones, preamps, amplifiers, and even the electronic self noise of recording devices. With 770dB of headroom above 0dBfs and 758dB of range below 0dBfs. Each amplitude increase is represented by a more scientific way of writing very large numbers compared to a traditional integer number.

It’s two main advantages (with caveats) compared to fixed point integers are: 

• Firstly that it records data above 0dbfs. Clipping your audio file is now impossible (but you can still clip the microphone, preamp and other stages in the chain).

• Secondly that it retains resolution at extremely low levels and doesn’t introduce any noise in the encoding of output voltages into a digital format (you are, at a very minimal level, still recording the self noise of the hardware matrix you record on and of course the ambient noise floor of the environment you exist in).

Key points directly from various writings on the topic

Sound Devices have been publishing blog posts about 32-bit float workflow since 2019. One of these was a comparison between the restoration of audio from 24 and 32 bit recordings. In this extreme example of restoring 100dB of gain, the 24 bit recording clearly fails to deliver enough useful information to make a presentable restoration. The reason for this is mostly because of the inherent problems of digitising audio using fixed integer values. We hit the noise floor and effectively hit the levels of imprecision at this incredibly low level of -100dB below full scale. I do think this is a very extreme case and you would very rarely come across a situation where 100dB of gain is actually required. In typical recording situations I think you would only realistically expect a swing of about 50dB in a scene. 

This is reflected in the other article that Sound Devices did, recording a spray and polish on the side of a muscle car, which is then revved hard. The recording clips 0dBfs by quite some margin, but the 32 bit float recording is “just” pulled back 22dB and we see the full range of the recording clean and clear. Here are links to two of those articles:

- 32-Bit Float Workflow: https://www.sounddevices.com/the-32-bit-workflow/

- Low Level Signals: 32-bit float vs 24-bit: https://www.sounddevices.com/low-signal-32-bit-float/

Sound Devices have written a number of articles on the topic now and they are all well written and worth understanding as much as possible. Some of those also include downloadable audio examples of the effects they are describing.

Conclusions

What I've gathered from everything that I've read since 32-bit float started being debated is that in essence 32-bit is a technically superior format, which is forgiving to errors in the basic gain staging of recordings, and in extremely dynamic situations can offer a way to deliver a recording true to the real dynamic range of the scene. In a discussion with a fellow recordist I joked that it would be perfect for a situation where a person is whispering and then a car explodes, without losing clarity in either the very low end of the dynamic range or clipping any of the loudest sounds. Your only limitation now is the dynamic range of microphones.

The file sizes are larger, but still tiny compared to 4K video formats typically used in modern video production. The only true downside is still (somehow) compatibility. Avid need to get their shit together and get 32-bit float working properly in Media Composer. At this point it's embarrassing to them and any editors locked into their ecosystem that it doesn't work. I implore everyone to look at different solutions than Avid products. Reaper and Resolve are really easy to use and don't take long to learn or even master.