Artificial Intelligence in Music: the last wall of the castle

Just a note to readers coming here for the experience of literary poetry combined with the original music stuff we do – I’m still doing some “summer vacation” writing that breaks from that form this month. This post does deal with music – if from another angle – and I expect to fully return to our traditional presentations this Fall.

So, I’m at my frequent breakfast place on a fine August morning that has not yet reached the AQI-alert level of smoke. In an unplanned coincidence, Glenn walks in. We’d talked last week about, of all things, Herb Alpert, and his early 1960’s instrumental hits, particularly “A Taste of Honey”  which was a chart topper in our youth.*

Glenn has some Herb Alpert and the Tijuana Brass CDs, but like many he’s as likely to have a CD player as he is to have a way to play 78 rpm shellac records. He’s been trying to get their music onto his new Mac Mini, but his old USB Apple Super Drive won’t recognize a music disk.**  Somehow (likely my current preoccupation with finally writing about it) we got to talking about AI. I mentioned that I’ve been struggling to use my collection of Virtual Instruments (VIs) to realize recordings with brass instruments that capture the full level of articulation the real thing can produce.

We talked a little bit about the various ways these instruments can be controlled: little plastic keyboards, various guitar pickup schemes, even wind controllers. Glenn has a bit of engineering background – this had (I hoped) some mutual interest.

I have little or no guilt in using VIs here for the Parlando Project. Not only is a VI grand piano highly affordable, it takes up no space, requires no fancy mic’ing, and produces a pleasing sound. Given my musical eclecticism, I think of how much more cluttered my studio space would be if I continued to collect odd instruments that I would experiment with to add unusual colors to pieces. And though I can’t actually play a real cello or violin, I can use a MIDI guitar controller to add those sounds. I’m grateful for those options for realizing my music.

Then I told Glenn about the Mellotron – a pioneering virtual instrument before such a thing had a name and acronym. Rather than hard drive files containing databases of digital recordings of actual instruments playing a range of notes in different articulations like one of my computer VIs, this primitive mid-20th century machine used strips of analog tape recordings of an instrument playing a single note for each tape strip. When professional musicians (among them: The Beatles, the Moody Blues, King Crimson, The Zombies) started to use the Mellotron, some objected: could the Mellotron put real musicians out of work? When the Beatles and their producer George Martin wanted a high trumpet part on “Penny Lane,”  a real musician was contracted for and played that difficult and memorable part. But flip the “Penny Lane” 45 RPM record over and on “Strawberry Fields Forever”  Paul McCartney pressed a Mellotron’s keys to produce an eerie flute sound. Listening closely, it wasn’t quite like a real musician blowing into a real flute. It was maybe 80% there – but if it sounded a little fake to a discerning ear, one might think it was still an interesting sound, whatever its level of verisimilitude. But imagine you’re a flutist in 1967 – the Beatles could certainly afford to pay for your services. Though bands moved on to use more complex synthesizers and other devices, real instruments still retained a level of preference when their fully-authentic sound was called for.

Could I pay or otherwise record real musicians instead of using my computer VIs? It’s hard for me to imagine a cello or violin player that would accept my chaotic and self-imposed quick-turnaround schedule, naïve/inconsistent musicianship, my shifting moods, and my no-revenue-project budget.

In my defense, this human being may well be playing the instruments,  just as I play guitar: this note, here, this loud, this long. Other times I’m scoring the music the VIs play, writing or modifying the MIDI event data rather than on a music-staff leger.

Still, there are some gray (or even darker) areas. For me, that started with using arpeggiators: ways to tell a computer you want it to take a chord and play the notes within it in a rhythmic series. I can tell it what note-length to use, something about the order of the notes, but the precision is then all the computer’s – and arpeggiators will have presets to suggest, and I might agree to one. Numerically quantifying the level of plausibility of my own work is problematic, but VI technology is such that even with my limited musical-instrument-operator skills, I may approach 90% there – but my musicianship, with its intents, and also it’s limits, is still involved. I can’t help but think my brass VIs sound badly because they are so far from the families of instruments I have played in “the real world.”

But a greater temptation arrives: more sophisticated computer “players” that take a chord sequence and duration I supply – from composition or by my playing something – and augment them by playing those cadences musically in a style it supplies and I consent to. These “players” have multiple adjustments, I can (and often do) modify what they supply as defaults, but this further development bothers me. Am I still the composer? In a human-musician world the answer would be clear by well-established tradition: yes, they’d say, I’m still the composer. Professional musicians, working before computer algorithms, have long supplied “feels,” timbres, expression, and entire decorative lines. They might even revoice the chords or play extended harmonies. They will do all that (or more, or better) than my computer does for me. So why do I feel bad when I ask my computer to do this? Well, there’s the impersonality to it. I’ve worked with others who’ve made important musical contributions to work I’ve originated, and that doesn’t feel the same. While I think I would be problematic to impose this on human musicians for the rewards I can offer, there’s more to it than not offering them that opportunity. I can’t help but think I’m cheating, that these realizations are fraudulent.

Yet guilt hasn’t stopped me from using these computer functions, and you’ve heard some of the parts they’ve played sometimes in Parlando Project recordings. The term artificial intelligence is elastic, it’s become a marketing buzz-word, but these enhanced arpeggiators and play-with-this-feel-or-articulation variations could fairly be called AI – even when the same musical piece has my vocals of I-hope-for subjective-quality or my it’s-supposed-to-sound-like-that guitar playing.

That said, over breakfast, I tell Glenn about how far AI music generation has come in the past few months. Just by entering a prompt or making a menu selection, often made up of generalized summary words for genre or playing style, one can create an entire song including vocals and all the musical accompaniment. Earlier in this decade the results would’ve been overly simple or subject to embarrassing defects. Now, the results easily pass the “Turing Test” for casual listeners. If the Mellotron flute is 80% there, and my best VI violin might be 90% there, these entirely machine-generated songs are about 95% there in verisimilitude. Sure, human musicians, real composers, even avid music listeners, are forever aiming for that extra 5% of skill, originality, and listener appeal; but when I listen to these productions which can be produced endlessly in minutes of hands-off computation time, the “tells” are the thoroughly AI songs meh obsequiousness to genre musical tropes and the slight artificiality of the machine-made vocalists. And that’s a problem. Centuries of musical theorists from the days of music theory treatises written with a quill, and onward to the accretion of hardened commercial songwriting craft, have supplied all the steps in ink-stained longhand to create a coherent musical structure with predictable effects. The computer coders only have to apply a light dressing of adaptation to transfer this consensus for robotic mass-duplication. The singers would still have remained a challenge – except by a fateful choice: popular music has increasingly prized machine-aided polishing of human voices to remove the inexactness they are prone too. Ironically, what could have been the last rampart to be surmounted by AI was dismantled by meticulous vocal production and ubiquitous auto-tune before the tech-bro Visigoths arrived.

I said to Glenn over breakfast in the café “Here we are talking about a popular song released 60 years ago, one we both still remember. ‘A Taste of Honey’  didn’t have any vocals, and now AI could easily produce an entire album of other instrumental songs to surround it – and even listening carefully, I’m not sure we could tell AI from human-written and realized musical pieces.”

This is not a theoretical exercise. Streaming platforms and playlists care even less than casual music listeners about AI content standing in for human work. In some genres, the algorithm that supplies your next song playing may already be a robot suggesting robots playing robot-composed imitations of human music. The only thing holding off an overwhelming onslaught of AI slop is that we, the audience, are still invested in the erotic worship of flesh-and-blood young performers and some residual romantic veneration of the human artist. Those things may be illusionary, but even if so, those things may be our defense. Do I have any other hope to offer? Yes, there’s something else, that comes next post.

This is the author of the play “A Taste of Honey” for which
the tune was composed. Her play frankly portrays a whole range of working-class situations in ‘50’s Britain. A teenager when she wrote her play, she was 21-years-old when this cheeky interviewer interrogated her. What admireable self-confidence!

.

*As vividly as I remembered the song, I knew nothing about its origin – and while I could distinctly recall the musical sound of Alpert’s recording in my head (trumpet, trombone, and that beating drum) I also heard in my mind vocals and a crooner singing. I tried to find the version with the sung lyrics I was remembering. I likely had heard the (somewhat unlikely) version of “A Taste of Honey”  done on the Beatles’ earliest LP, but I don’t think it was that one I was hearing in head.

**If you still own that ancient Apple artifact, the external Super Drive CD/DVD drive, you should know that it won’t work unless connected directly to one of your Mac’s USB ports. Even deluxe powered USB hubs or docks won’t work–  the drive will seem completely dead when connected through them.

5 thoughts on “Artificial Intelligence in Music: the last wall of the castle

  1. This is a thought provoking post, Frank. I’m not a musician nor any kind of performer, really, but it seems to me that a musician (or even a person with a really good ear) might be able to distinguish between computer generated music and music performed by humans. Does it contain “errors” or idiosyncratic embellishments? Probably a human. (I’m suggesting that it may be easier to tell what is NOT from a computer.)

    There are and have been human performers and composers who have demanded perfection from their collaborators… Captain Beefheart and Raymond Scott are two that float to the top of my mind. Raymond Scott’s Wikipedia entry is especially worth looking at as he was an early developer of electronic music and instruments, partly for that desire for exactitude.

    Finally, music is a social activity, even if it is a computer file and a solitary listener. Otherwise it’s just another falling tree in the forest. Regardless of how the sounds are produced, if there is an audience for it then it’s music. (The audience’s enthusiastic reception is one of the reasons why I found the LYL Band’s alternative prom performance so delightful.)

    Liked by 1 person

  2. I still have a turntable and a couple of 78 shellacs, some 45s and 12 inch LPs ah, those were the days. I fell in love with the idea of mellotrons – some great work in the early days by several artists.

    Liked by 1 person

  3. While I have (and use) much more realistic Virtual Instruments with gigabytes of digital samples of every note and playing technique for a real instrument, the approximation of the Mellotron still has so much quirky charm. I even sort of like it when the Mellotron Virtual Instruments retain the limitation of the original and just cut off the sound after the key has been down for more than 8 seconds because each real Mellotron note was a strip of tape, not a loop. When it came to the end of the tape strip things had to mechanically snap back before that note could be sounded again. Similarly, the wheeziness of some pump or bellows reed organs or such like where you get a fade as the bellows air chamber depletes.

    Right now AI doesn’t understand that kind of thing very well, the glitches, the limitations left in, the man’s reach exceeds his grasp stuff that music can tolerate.

    Like

Leave a comment