Song for Jane
Meet Jane
Let’s say Jane is 33, and she sings beautifully. She sings like she means it. Her voice is deep and smooth, even when she hits the high notes. Her range is solid, her mix is balanced. She’s sung in every neighbourhood choir since childhood. She sings at weddings. She sings at karaoke nights. She brings her guitar to every community picnic. Without fail, someone new always asks, glass in hand: “Have you ever thought about going pro?”
No, a singing career never quite happened for Jane. She auditioned in college. But like thousands before her, Jane hit the wall of statistics. She may have a gorgeous voice and memorise all of Adele. But how many girls show up regularly for a spot in a band? And how many of those spots regularly open? These were Jane's chances, because
Jane does not write her own songs.
She has to borrow. By her third garage cover band, she’d learned the lesson:
If you never find your Jim Steinman, you’re stuck in a waiting room. Holding a number. Praying, no one before you picked the same song.
Jane went to college for something unrelated to music. She found her first dead-end job. Then a better one. Then a husband. Then find herself divorced. Let’s say, she works in a bookstore and raises her preschool daughter. Still killing it at karaoke. And maybe, just maybe, Jane still entertains the idea of doing her own thing.
I know a Jane in real life. A few actually. Sitting next to them at Friday karaoke, I couldn’t help but wonder: What can AI actually do for someone like Jane?
There’s nothing written by her to begin with, no lyrics — she’s not sitting on a cache of training data. And yet, she needs something personal.
Subscribe
What are Jane’s options?
Prompting
Jane first heard about AI in 2022 — impossible to miss when GPT-3.5 had entered the mainstream. Suddenly AI is going to steal our jobs… and data… and grandma’s apple pie recipe. But right now AI also writes. Jane opened up a prompt window and typed something like: “Write a folk song about loneliness.”
And she probably got something like:
“The stars are cold, the night is blue / My heart is gone, away with you.”
Yes, it is GPT-3.5, in its most earnest attempt.
I imagine Jane read it, closed the window, and kept ChatGPT for writing customer service emails. Tasks that require zero risk of self-expression. I sympathise though. My first AI-generated LinkedIn post — a new position announcement:
“Inna brings a vibrant passion for finance and software, melding it beautifully with a profound knowledge of Data Analytics.”
Makes me cringe to this very day.
AI-Powered Tools
At some point, Jane found her way to songwriting tools. In recent years, there came a number of beautiful AI-powered assistants, and their services expanding constantly. Jane can pick a genre, mood, a few keywords — maybe even drop in an artist name. She could press “generate,” read the lyrics, press it again until the phrasing “feels” closer to what she want. The templated generation approach, however, comes with limitations. You can’t easily:
Provide personal writing samples to adapt the tone;
Edit and re-post fragments for iterative refinement;
Guide the model to follow a specific song structure…
Under the hood, most of these tools are wrappers around LLMs. They typically rely on zero-shot prompting or predefined prompt-engineering templates. Some incorporate fine-tuning or parameter-efficient adaptation methods to narrow the model’s behaviour toward desired outputs.
Still, I encourage you to try them — they are genuinely fun! You might get a spark and find a line that clicks.
I apologise in advance if any of these point are already outdated. Given the pace of innovation, there may be a new unicorn launched yesterday, solving all of the above.
Custom Fine-Tuning
A model explicitly trained on a user’s data is the current standard for personalisation. It allows the model to internalise a specific style, tone, rhythm, fingerprints... Tools like Hugging Face’s AutoTrain have gone a long way toward democratising transfer learning.
But fine-tuning still requires:
At least some degree of technical proficiency;
Computational resources, ideally GPUs or high RAM for CPU-based fine-tuning;
And most painfully: a sizable, high-quality training dataset.
Fine-tuning pipelines assume you already have a substantial, meaningful text corpus. But how many singers — professional — have thousands of training-grade samples?
In Jane’s case, this is where the dream collapses.
Jane meets LoRa
LoRA — Low-Rank Adaptation — is one of adapter-based methods. Thay allow LLMs to learn without full/partial retraining. LoRA inserts small trainable modules into the attention layers — and only those get updated.
Sounds like a path forward for Jane.
I built a small project:
👉 https://github.com/InnaVays/song-for-jane
It took about about 15 hours (thank you, Copilot) and one slim book of songs — Songs of the West (thank you, Gutenberg archive). A thousand stanzas of public domain folk lyrics — clean and stylistically consistent. For the foundation I used one of Microsoft’s lightweight model, well-suited for experiments and manageable even on CPU.
The results were surprisingly delightful:
Prompt:
“Write a folk-style song about a church bell and a goat.”
Output:
“Why, my sweet little goat, I'll never see thee again,
I'll marry my true love as soon as I can,
And my true love will be as old as the hills
And the church bells will ring through the fields”
I can almost see villagers singing it on the way home from… somewhere.
Next: Inventing Jane
Now we have a starting point — a lightweight, reproducible training pipeline and a small model that can echo a single voice. Maybe Jane and LoRA could become friends?
But this is also where the real challenge begins. If the model must truly reflect Jane, we’ll need to create something that has never yet existed — synthetic data from Jane, when Jane has never written a line.
If you’ve read this far, I thank you. This is the very first post in Make It Personal — my personal attempt not to drown in generic AI. As someone wise once said: the only job of the first step is simply to be made. Though, it might’ve been ChatGPT.
Thanks for reading Make It Personal!