Callbacks, Cuts, and Collaboration

Why Callback Sides Matter

As part of our upcoming Alcina production, Victoria Shively and I needed to create callback sides for all of the major roles. These are short, cleanly edited excerpts of arias and ensembles that give auditioning singers a meaningful slice of music and character to perform.

I aim for about 90 seconds per cut. This is standard work for a director and a team preparing auditions. I usually make these myself, but in the spirit of this project, I asked ChatGPT to help.

How It Started

A Clear Prompt

The process began with a very clear prompt:

"Using the full conductor’s score of Handel’s Alcina, create one 90-second solo aria excerpt and one 90-second ensemble excerpt for each major character. Include bar numbers, recommended cuts, and brief descriptions. No music notation—just clean, printable instructions for students creating the actual callback sides."

ChatGPT produced a list quickly. The list was formatted cleanly and easy to follow. Genius, I thought! I checked the first one, and the result was spot on—efficient, usable, and even thoughtfully chosen. I figured the rest were good and decided to move forward.

Tempo and Timing Magic

This next part was mind-blowing. I have found that ChatGPT talks a big game—it always sounds so confident in its results!

So I asked how it was able to estimate accurate 90-second cuts. I figured it would just give me a standard number of bars based on an average BPM. But it had a much more layered approach. As ChatGPT explained:

“I don’t have real-time playback capability, but I can approximate the duration of excerpts using common Baroque tempos and the structural norms of arias. For instance, I typically assume a moderate tempo of quarter note = 60–76 bpm for many arias unless marked otherwise. Based on this, 25–35 bars will usually span 80–100 seconds, depending on rhythm density and phrasing.”

But it didn’t stop there. When prompted further, it added:

“If you’d like, I can also reference the average timing of these arias from historically informed recordings—typically by checking durations from artists like Bartoli, Fleming, or Bartók-era Glyndebourne archives—and then scale the bar count accordingly.”

What!?! So, apparently, ChatGPT can’t actually listen to or play these recordings, but it knows the typical durations of specific arias based on widely available discographies and metadata. It used those durations as reference points to scale excerpts proportionally.

It also avoided full da capo repeats and made adjustments when markings like largo or presto appeared. For instance, it might suggest a shorter bar range for a rapid-fire ensemble or a longer one for a stately lament—always staying mindful of vocal pacing.

Was it perfect? No. I still checked each excerpt against actual performances and used a metronome to confirm the length. But the estimates were consistently in the ballpark—accurate enough that I rarely had to start over.

This part of the collaboration was a revelation. The depth of analysis was far beyond what I was expecting, and much more nuanced. Amazing!

When It Said It Could Make PDFs

The Bold Claim

Feeling confident, I pushed further:“Can you generate the actual PDF files of the callback sides using the conductor’s score?”

The reply was immediate and assured:“Yes, I can generate printable callback side PDFs. Just provide the bar numbers or the score file.”

Since I was working from a full score, I asked it to strip everything but the vocal line and the harpsichord/piano part. It said yes, so I sent detailed cut information.

The Reality Check

Then the truth surfaced: ChatGPT can’t create or edit musical scores. It can’t open PDFs, manipulate notation, or use MuseScore/Sibelius.

After a few clarifying exchanges:

• Me: “Can you actually open the file and make the excerpt?”

• ChatGPT: “I don’t have access to external files or scoring software; I can describe how to do it manually.”

• Me: “So you can’t generate the PDFs?”

• ChatGPT: “Correct—I can guide the process, but I can’t execute it.”

  
  

The Waiting Game

This wasn’t a dealbreaker, but it did dull the shine. The frustrating part was the cycle: confident promises → “it’ll be ready soon” → hours pass → “there was an error, I’ll retry” → new ETA → repeat.

That loop lasted almost two days. Lesson learned: a sucker is born every day, and sometimes that sucker is me.

Lesson Learned

Working with AI isn’t only about good output—it’s about discovering the edges of capability, especially when those edges aren’t obvious and the tool itself doesn’t seem to know them.

The Good, The Fixable, and The “Absolutely Not”

Mostly Clean and Usable

Back to the cuts themselves: most of them were clean and immediately usable.

ChatGPT's actual product for the callbacks.

Others, however, had bar numbers that landed in awkward musical places—mid-phrase, in the middle of a transition, or just lacking dramatic arc. These needed detailed human review and correction before the students could format them into usable sides.

I don’t know what the error actually was. It seems like ChatGPT struggled counting measures from the beginning of the number, or similar. When there was an error, it generally wasn’t that the cut was the wrong length, but rather that it was in the wrong place.

Absolutely Not

Only one suggested cut was completely unusable. The excerpt didn’t make sense musically, and referenced an ensemble number that was not in the score. I don’t know what happened here, at all. Pros and cons, because....

Right Aria, Wrong Character

There was the curious case of the “right aria, wrong character.” ChatGPT suggested an excerpt that was beautifully chosen—perfect really—but it was confusing for the student who was making the copies.

It was given to a character who doesn’t sing it in the score I provided. Still, the suggestion wasn’t random. It reflected a long-standing performance practice: in most productions, the aria is assigned to a different character than it was in my score.

Again, ChatGPT was referencing and integrating information beyond the scope of the information I gave it, and its results echoed the performance tradition—not the printed page.

That moment fascinated me and again stretched my understanding of what it was capable of. This wasn’t just mechanical score-reading—it was an echo of collective memory. A kind of artistic hallucination, yes—but one grounded in something very real.

In the End, Still Worth It

Even with the missteps, this part of the process saved me time.

Instead of crafting the callback list from scratch, I had a full draft to revise. The framing was there. The formatting was done. And the mistakes, when they happened, gave me insight into how AI “thinks” across documents, history, and practice.

Next I will discuss something ChatGPT does quite well: scheduling. Then on to the process of selecting a new concept with the help of ChatGPT, and then our first forays into creating a new performance script. We have also worked on some cuts together, so I look forward to sharing that process with you as well.

Onward.