AI, DAWs & Audio 2

For perspective, let’s look at ChatGPT and its training process…

‘ChatGPT’ stands for ‘Chat Generative Pre-trained Transformer.’ ‘Chat’ refers to its primary function of engaging in conversational interactions; it is fundamentally a chatbot. ‘Generative’ refers to its ability to generate an output (in this case text responses) based on the input it receives. ‘Pre-trained’ means it has been trained on a vast set of relevant data prior to being fine-tuned for a specific task, enabling it to understand and produce human-like text. ‘Transformer’ refers to the underlying architecture used to process and generate language; the Transformer model is a popular framework in natural language processing. More than meets the eye, huh?

The term ‘ChatGPT’ therefore describes an AI model that’s designed to generate conversational text based on prior training and advanced language processing capabilities.

ChatGPT-3.5 was the first version to go public, in November 2022. It was trained using a Machine Learning technique called ‘supervised learning’, followed by ‘reinforcement learning’. It was pre-trained from vast datasets containing text from books, websites, and other publicly available sources – but its training was limited to information that existed up to September 2021, a date known as its ‘knowledge cut-off’. Initially, human trainers provided example conversations to guide its responses. Then, reinforcement learning with human feedback (RLHF) fine-tuned its performance, where human evaluators ranked its output and provided feedback to improve quality. This iterative process enabled ChatGPT-3.5 to generate coherent, contextually-relevant responses across diverse topics.

On March 5 2026 OpenAI released the latest versions of the ChatGPT ‘family’: ChatGPT-5.4 Thinking, ChatGPT-5.4 Pro and ChatGPT-5.4 Mini. These versions all have a knowledge cut-off of August 31 2025. To reduce hallucinations when formulating responses based on information beyond their knowledge cut-off [as discussed in the first instalment of this series], The ChatGPT-5.4 family relies on advanced web browsing techniques and native computer interaction (i.e. mimicking human behaviour when searching and checking information), allowing it to search the internet, read screenshot-based data (i.e. extract text from images), and interact with other software to fetch real-time information. Nonetheless, any results based on information beyond its knowledge cut-off should still be treated with caution because, just like when a human is searching for information, it’s possible to come up with two related data points with missing or conflicting information between them, and assume – or hallucinate – the most plausible connection.

[Scroll down to ‘Educating Leta’ for a real-world insight into AI testing and training.]

AI-ASSISTED DAWS FOR MIX ENGINEERS

Because this magazine is called AudioTechnology we’ll now move away from chatbots, data searches and hallucinations, and focus on an AI intended to exist within a DAW and assist with audio production duties. Many of the processes discussed below have equivalents for those working in video production, and those equivalents would apply to an AI-assisted NLE.

WORTH PAYING FOR

I’d pay for a pre-trained, validated and tested AI-assisted DAW that further trained and personalised itself on my session files, my working habits and my aesthetic, as described above. The first thing I’d train it to do is the grunt work I normally do before mixing, as follows…

1. Working Copy

Make a copy of the client’s original folder and send the original folder into cloud storage for safe keeping – thereby ensuring the client’s original files are not affected by any of the following steps.

2. Metadata Correction

Analyse each audio file’s metadata and repair/correct if necessary to prevent the inexplicable “I don’t understand, it worked perfectly yesterday” problems and the bounce issues that inevitably occur when one or more files in a session have metadata problems such as incorrect file sizes and missing EOF conditions. I consider this to be a vital first step when receiving files from someone else to mix – especially if the client has bounced those files out of a DAW after doing all of their desired edits, auto-tuning, and so on.

3. -20LUFS Normalisation

Adjust each file’s loudness to sit around -20LUFS unless this causes its peak level to exceed -1dBTP, in which case reduce the level so it does not exceed -1dBTP and make a note of the applied level reduction. This is a simple analysis and gain process that can be done by the AI before the file is placed into the session. The benefits of this intelligent normalisation are explained in step eight.

[I currently use Zynaptiq’s Myriad batch processing app for steps two and three above. (‘Batch processing’ simply means all the files can be loaded in at once and have the same processing to all of them, or selected groups, at the press of a button.) For step two it identifies any metadata issues and offers a choice of ways it can correct them, along with the pros and cons of each approach. This is a purely objective process that an AI could do while also implementing ‘best choice’ solutions based on the context and type of problem, leaving less of the decision making for me. For step three Myriad lists the current LUFS and TP values of each file, and can simultaneously normalise all files (or a selection of files) to a given loudness value. If the supplied files’ LUFS and TP levels vary wildly between each other, I use a spreadsheet that calculates how much gain I can safely apply to each file to keep them all files at the same LUFS level before loading them into the session file. This is objective ‘grunt work’ that an AI can perform and follow through as described in step eight.]

4. Create The Session

Create a session and load the checked and adjusted files from the previous steps in the order I’d usually place them, grouping related sounds together (e.g. bass DI channel next to bass mic channel). Name all tracks and channel strips according to their file names. I can abbreviate the names to something more intuitive later if necessary, but if there’s something wrong or dubious with a file I want to know at a glance the name it was supplied with to ease communications with the client.

5. Plug-Ins & Sends

The AI will insert the plug-ins I typically start with on each channel strip: one corrective EQ followed by one corrective compressor followed by one enhancing EQ and one integrating EQ, as described in the fourth and fifth instalments of my Mixing With Headphones series. All plug-ins are the same on every channel, just like in the days of analogue consoles where every channel strip had the same EQ and the same dynamics processing (if it had any dynamics processing at all) and engineers got on with mixing the music rather than pondering which plug-ins would be the best choice, because there was no choice – except maybe two or three processors in the effects rack that had to be connected via a patch bay and cables. Note that all plug-ins are bypassed at this point. The AI will enable a channel’s corrective EQ if required in step six, and will enable a channel’s corrective compression if required in step eight.

Also during this stage the AI will set up my usual spatial processing configurations: creating additional channels into the mix for each of my spatial processors, setting up auxiliary sends to access them, and setting ‘start point’ predelays, reverberation times and similar parameters based on the music’s tempo (as explained in the sixth instalment of my Mixing With Headphones series).

6. De-Guffing With HPF

The AI will find the lowest frequency of musical value in each track (based on the frequencies of the notes in the music’s scale) and set a high pass filter at the appropriate frequency and slope on each channel’s corrective EQ to remove any low frequency guff the AI assumes will have no musical value.

I’ll double-check the AI’s choices as I work through the mix – but I don’t want that guff getting in the way while I’m making EQ decisions. I also don’t want inaudible, unnecessary or non-musical subsonics influencing the behaviour of dynamics processors or forcing my monitor speakers to reproduce frequencies they cannot without compromising their performance. I’d rather start with that low frequency guff removed and ease it back in if desired.

7. Temporary Mutes

Apply mute automation to any non-musical parts at the start, end and during tracks (count-ins, chat between musicians, etc.). I don’t want these things deleted, just temporarily muted, because sometimes they add value to the mix from a performance point of view. I’ll be checking their potential value to the mix later.

8. Clip Prevention

Tracks that had their loudness level reduced below -20LUFS to prevent exceeding -1dBTP (as described in step three above) will be re-assessed in case their peak levels have decreased after the removal of unnecessary LF energy (step six). Tracks that still have excessive peaks will have those peaks reduced by the corrective compressor in the channel strip. The AI will apply ‘low threshold/low ratio’ compression using objective mathematics (as demonstrated in the fifth instalment of my Mixing With Headphones series) to prevent those tracks from clipping – regardless of whether the result sounds good or not. Making that corrective compression sound good is a subjective thing I’ll do later; all the AI has done is ensure the sound is at a usable level to start the mix without clipping. Every channel that has this type of corrective compression applied by the AI will have a red fader top, indicating that I should check it. It’s also possible that the corrective compression has resulted in a higher LUFS level, which is addressed in the next step.

9. Fader Levels

Due to the steps above I know that all tracks have approximately the same loudness level of -20LUFS – other than those with excessive peak levels (which were made lower in level in step three, but might now be perceived as louder due to the corrective compression applied in step eight) and those that had significant subsonics removed in step six (which may now have lower overall levels).

This final step in the set-up puts all channel faders to the same level, determined by the AI to provide an average of 20dB of headroom above -20LUFS through each channel strip. The channels are then muted. For the sake of this discussion we’ll call this the ‘norm’ position. Tracks that have altered levels due to corrective compression (step eight) and/or corrective EQ (step six) will have their levels compensated with the plug-in’s output level control if available and sufficient. If not, their levels will be compensated by the fader and indicated with a red fader top (if not already).

10. Ready For Mixing

Most of my tracks are now sitting around -20LUFS and have about 20dB of headroom above that nominal level in their channel strips. This creates an equivalent to the traditional analogue studios of the past, where every track was recorded at a nominal level of 0dBVU (and therefore at approximately the same perceived level) and every channel strip had at least 20dB of headroom above that nominal level of 0dBVU. This is a starting configuration I’m very comfortable with because it’s what I learnt to record and mix in. It’s also a good starting point for the AI to create its own balances from if required (more about those later), and is invaluable when using plug-ins that emulate vintage compressors and limiters because those devices’ thresholds were intended for use in analogue systems that had similar signal characteristics to those created through the steps above.

Sleeves Up & Get Working

The steps above represent the worse-case process I go through when preparing a mixing session using files I did not record myself, although some of the steps have been both simplified and taken further due to the AI’s capabilities. Now it’s time to check the AI’s decisions by briefly soloing each track before rolling up my sleeves and getting to work.

Perhaps the AI set a channel’s HPF too high (step six), and lowering it an octave adds inexplicable ‘weight’ to a sound despite having no identifiable musical value. Perhaps a sound that’s been compressed/limited due to excessive dynamics (step eight) was never intended to be very loud in the mix and therefore doesn’t need any compression or peak limiting. Perhaps a 1/64th note pre-delay for reverberation is too short for any of the spaces I want to create for the mix (step five). Perhaps I like the ‘up’ vibe of the muted chit-chat between the drummer and bass player at the end (step seven) and want to include it in the fade out.

I can change any of the AI’s decisions at any time, but, assuming all is well, I’ll be starting my mix from a well-informed position based on my own preferences…

From a stability point of view, I know that all of the tracks have had their metadata checked and corrected, so I won’t expect any last-minute surprises during bouncing or saving.

Tonally, I know that most of my tracks will be sitting around the same point on the Equal Loudness Contours when heard on the mix bus, ensuring my initial corrective EQ and enhancing EQ decisions are made in the same tonal context – and that will remain true until I start pushing faders around, at which point I might tweak my enhancing EQ while also using my integrating EQ to help sounds sit alongside each other in the mix (as demonstrated in the fourth instalment of my Mixing With Headphones series).

Dynamically, I know that channels with red fader tops need my attention. In most cases the red fader top indicates the AI’s corrective compression has been applied in a purely objective manner to prevent clipping (as described in step eight) and will probably need altering to achieve a musically acceptable result – considering a) the dynamics of the track and its role in the music, b) the dynamic perspective I’m working in for the mix, and c) the dynamic limitations of the release medium. There’s no point in allowing peak levels in a mix that I know will be problematic downstream for mastering and/or distribution. Prevention is better than cure. If I’ve changed the AI’s objective settings on a corrective compressor the red fader top will return to its normal colour. In the remaining cases it means the corrective compression or corrective EQ plug-in used on a channel strip had insufficient or no output level control, and the signal level was balanced against the others using the fader (as explained in step nine).

Spatially, I know that my preferred reverbs and delays have been set up on auxiliary sends and are returned into the mix via their own channel strips – meaning they can have EQ and other plug-ins applied to them, and each effect can be sent into another to create layers (more about that in the sixth instalment of my Mixing With Headphones series). Each channel’s auxiliary sends have been set to post-fade and their send levels reduced to -∞dB (i.e. off). Each auxiliary send’s master level has been set to 0dB, and each spatial effects’ fader to the mix bus has also been set to 0dB. To apply a spatial effect to a track I simply need to turn up the appropriate auxiliary send on that track’s channel strip – the rest of the effect’s signal path is open and ready to go.

With coffee in hand, butt on seat and headphones on head, I’m ready to start mixing the session. I’m using a remotely backed-up set of files with no metadata errors, no excessive peaks or subsonic surprises, and no complex patching headaches – yet all I’ve done is point the AI-assisted DAW to the folder containing the client’s original files, told it to prepare a mixing session, and wandered off to get my morning coffee (a necessary part of my daily ‘becoming human’ routine). The AI-assisted DAW did all of the time-consuming grunt work and brain-draining screwdriver work for me, in the way that I like it done. I can now audition the tracks one at a time, in groups, or altogether – whatever feels like the right way to start the mix.

The steps described above can sometimes take me up to an hour to complete and leave me mentally exhausted before I’ve even started mixing. An AI-assisted DAW could probably do it all in five minutes – especially if the AI’s hardware was built into the device running the DAW (as it is with the Apple Intelligence and Logic Pro combo) or is directly connected to the device rather than using a wireless internet connection. That’s a potential saving of an hour at the start of a session. I’d still charge the client for the agreed time/price, but I’d use that time saved by the AI, along with my fresh mind, to get everything ‘just right’ rather than ‘good enough’, and ultimately create a better mix that serves the music (as discussed below).

FURTHER CAPABILITIES

The 10 steps described above are simple tasks for an on-device AI, and could be set into motion by selecting a menu item (e.g. ‘Set Up Session’), navigating to the folder that contains the client’s files, and clicking on ‘Start’. The mix engineer can then make a coffee, have a shower, check the socials or whatever while the AI-assisted DAW does the grunt work and screwdriver work of preparing a session for mixing.

That’s great for setting up, but what about during the mix?

Apart from asking the AI to create a balance to start a mix from or to send as a preview to the client (we’ll discuss the differences between a ‘balance’ and a ‘mix’ shortly), here’s a brief wish-list of useful tools I’d be asking for. No doubt you can think of more…

I’d like to select a track and click on a menu item ( ‘Auto Adjust’) that tells the AI to provide a basic tonal and dynamic adjustment of the track. It’s an advanced version of steps six and eight above, but refers to its pre-trained knowledge of spectral references and dynamic references relevant to the sound source and the genre. An ‘intensity’ control would allow the processing to be increased or decreased with a single slider – similar to the auto-adjust feature found in photo editing apps, where increasing or decreasing the overall effect creates individually scaled changes across many parameters at once.

I’d like to select two or more tracks that are competing with each other in the mix despite my best efforts, and choose a menu item (‘Separation’) that tells the AI to attempt a solution using the current plug-ins and settings as prompts for the outcome I’m aiming for. I can undo it if I don’t like it, alter its strength if the result is too extreme or not extreme enough, or request another solution.

Conversely I’d like to select two or more tracks that feel too separated and choose a menu item (‘Ensemble’) that tells the AI to attempt a solution that reduces the separation between them (e.g. making a close-miked drum kit sound like a single instrument – which it is), again using the current plug-ins and settings as prompts for the outcome I’m aiming for. I can undo it if I don’t like it, alter its strength if the result is too extreme or not extreme enough, or request another solution.

Similarly, if I can’t get a track into the correct dynamic perspective in the most transparent way possible, I’d like to select the track and choose a menu item (‘Match Dynamics’) that tells the AI to attempt a solution by using the existing plug-ins and settings as prompts for the musical dynamic range I’m aiming for (the concept of musical dynamic range is explained in the fifth instalment of my Mixing With Headphones series). The end result might not sound as good as I’d want it to, in which case I was probably striving for an unrealistic outcome to begin with. Sometimes it’s impossible to make a sound with wide musical dynamic range sit comfortably in a mix with a narrow musical dynamic range while maintaining the same tonal qualities. As Roger Miller wisely advised, “You can’t roller-skate in a buffalo herd.”

All of the things mentioned so far in this article are better for me, and they’re better for the client. Everyone wins because the AI has been trained on my working methods and my understanding of the work. It’s doing the grunt work and the objective calculations much faster than I could, and it’s leaving the subjective decisions, creative ideas and time savings to me. Everything is as it should be in the human/AI relationship – unless you work as an assistant engineer, in which case you’re being shuffled aboard the same lifeboat as the readers and compositors mentioned in the first instalment of this series. More about that problem in the following instalments…

SERVING THE MUSIC

I’m confident that my mixes will always be musically superior than any mixes that today’s Narrow AI can create. Why such confidence? Let’s start at the very beginning, a very good place to start…

In the earliest days of my audio engineering career I learnt that musicians would always choose a bad recording of a good performance over a good recording of a bad performance. The bad recording obviously had some imperfections – maybe too much tape hiss, or a moment of clipping on one of 24 tracks – but if it was a good performance the musicians would always choose to move on rather than record it again at the whim of a technical perfectionist like me. Perhaps it was a time limitation, perhaps they couldn’t hear the technical problem, or perhaps they quietly knew it was the best performance that musician was ever likely to give. Whatever the case, I had no choice but to live with this technical problem, and eventually it became the loudest thing I’d hear in the recording. It haunted me. There was a sense of dread as it approached, a sense of regret as it passed, and every time I’d tell myself, “If only I’d taken a moment longer…”

Knowing that a good performance would always beat a good recording pushed me towards refining my recording and mixing skills until they became fast and instinctive, which, along with heightened attitude maintenance (dealing with size 10 musicians wearing size 20 egos), allowed me to aim for the triple win of making good recordings and good mixes of good performances. In other words, using my skills to serve the music – which requires some inherently human sensitivities that current Narrow AI does not have the sensory inputs for. Let’s explore them…

The EEG headset shown above looks like a swimming cap covered in sensors. Each sensor is measuring tiny electrical impulses that occur at different points over the brain’s upper surface and pass through the skull, where they are measured by the sensors pressed against the scalp. These electrical impulses are caused by groups of neurons in the top few millimetres of the brain’s surface firing off tiny electrical discharges at about the same time. They repeat this behaviour on a regular basis, but the time intervals between impulses depends on the body’s current physiological needs and also in reaction to external stimuli. There’s an entire science dedicated to interpreting these electrical impulses and what they mean. Thankfully we don’t need to go very deep into that science. For our purposes we only need to know three things about these electrical impulses: when they occur, where they occur, and how fast the EEG can respond to them.

When?

How often the electrical impulses occur tells us what ‘state’ or ‘behavioural mode’ the brain is in:

In Delta state the electrical impulses occur four times per second or less, which indicates deep sleep.

In Theta state the electrical impulses occur between four and seven times per second, indicating drowsiness.

In Alpha state the electrical impulses occur between eight and 12 times per second, indicating an awake but relaxed condition.

In Beta state the electrical impulses occur between 13 and 30 times per second, and indicate active mental engagement.

Where?

Where the electrical impulses occur also tells us something useful. Putting aside the left/right brain theories and focusing instead on emotional states, electrical impulses in the left frontal lobe are indicative of approach-related emotions such as happiness and joy, while electrical impulses in the right frontal lobe are indicative of withdrawal-related emotions such as dislike and fear.

How Fast?

The EEG’s sensors directly measure the electrical activity of neurons, so electrical spikes (0.02s to 0.07s) and impulses (0.07s to 0.2s) are measured as they happen, effectively at millisecond scales. There might be some latency as the EEG’s software processes and presents this information to the operator, but for all intents and purposes it is essentially instant – certainly fast enough to associate a momentary response with a momentary stimulus, such as a brief but exciting hook in a song…

A Penny For Your Thoughts

We don’t need a swimming cap full of sensors to measure the things described above. We can do it with just two sensors, or two small clusters of sensors, one on each side of the head. Damn! If only there was an everyday audio tool that allowed us to place a sensor on each side of a listener’s head without making the listener look or feel stupid – which might skew the results. We’d be able to bring EEG data into the mixing process – fine-tuning our mixes, and perhaps the composition itself, to maximise listener engagement…

The active noise-cancelling headphones shown above, Neurable AI’s MW75, contain EEG sensors within their ear pads (the light grey squares) and currently work with a mobile app to present EEG data related to focus, productivity and drowsiness. In terms of build quality, comfort and sound quality, reviewers compare them favourably with other premium headphones that offer active noise cancelling.

EEG headphones such as these could provide an AI-assisted DAW with the information needed to expose strengths and weaknesses in the mix, or even in the music itself. Correlating the EEG data with the mix would reveal parts with high engagement (Beta state), parts the listener likes or dislikes (frontal hemisphere asymmetry plus engagement), parts where the listener relaxes into the music (Alpha state), boring parts (Theta state), and parts that send the listener to sleep (Delta state). This data could be measured throughout the duration of the music, entered into the AI-assisted DAW, and displayed as a trace on the session’s timeline called ‘Engagement’ or similar. This information would provide the mix engineer and the AI with an excellent indication of where parts of the mix would benefit from being pushed harder or softer, and where the mix is losing the listener’s attention. From a mixing point of view, at least, this information could be used to optimise listener engagement. Perfect for algorithm-driven music distribution platforms such as Spotify, that rely heavily on engagement as an indicator of what to recommend to a given demographic.

Ideally the EEG data would be gathered over a number of listeners. Some basic processing (essentially a clever form of integration) would remove irrelevant responses that are unique to an individual listener, and, from there, an Engagement trace could be displayed on the DAW in parallel with the tracks or overlaying them. A convenient way to capture enough EEG data would be to put the finished mix into a smart phone (before it’s uploaded to any distribution platform), and pass the headphones around half a dozen like-minded friends. This is essentially the same as the market researchers’ ‘focus group’, which typically contains five or six participants – enough to gauge a general consensus.

Unfortunately, the EEG headphones shown above are large and bulky to carry around, while also being visually conspicuous and capable of ruining a time-consuming hair-do – meaning the listener might feel uncomfortable about putting them on, and that could skew the data as mentioned earlier. The EEG data gathering process would be a lot easier if all that was needed was a smart phone and a set of earbuds…

The product above is a pair of EEG earbuds and their related mobile app from Emotiv. The entire system fits in a pocket, the earbuds are sufficiently inconspicuous, and they won’t upset any hair-do. The operator (mix engineer or recording musician) could be watching the EEG data on their device’s screen in real-time, keeping an eye out for anomalies that have nothing to do with the music but coincide with events happening nearby – such as a spilt drink, a passing ambulance, or a thumbs-up from the hairdresser.

No matter how far-fetched this may seem, remember that market researchers have been using EEGs for years to determine if someone likes or dislikes something, regardless of their personal biases. It is, potentially, the ultimate measure of engagement… [Scroll down to ‘Other Uses of EEG’ more interesting applications of EEG and similar.]

In a world driven by analytics and engagement, an AI-assisted DAW with EEG feedback allows everything to be tweaked for optimum listener engagement: including the composition itself and every sound and instrument, all the way through to mixing, mastering and promotion. These capabilities will become particularly interesting in later instalments of this series, when we look at AI-assisted DAWs for recording musicians…

[The grey-coloured text within this article was written by AI. Each section took less than 10 seconds to prompt, create and paste into this document, and required no preliminary research on my behalf. I did some fact-checking and hallucination-checking afterwards, then edited the text just as I would do with anybody else’s writing to bring it in line with AudioTechnology’s ‘voice’. In each case I was exploiting AI’s strengths (information gathering) and avoiding its weaknesses (judgements and opinions).]