THE THING WE BUILT:

One Song Studio

Maybe you’ve seen some fancy tools that helps you generate a song. Type out a prompt and 🪄Voilà! 🎵: New Song. (But it’s probably boring or terrible.)

I’ve played with these a bit. You get to something almost decent... if you describe everything in great detail. There’s one problem though: it’s joyless.

I’ve experimented with a lot of AI tools. Some of these new tools are a dream for a tinkerer. Here’s what my goal isn’t: use AI to create art so I can do more spreadsheets. Outsourcing songwriting kind of sucks.

In Jeff Tweedy’s book How to Write One Song, he gives you tools and exercises to write the one song. Some are as simple as letting go of the fear of failure (maybe the true mantra of this newsletter). Others are practical exercises that chip away at the myth of writer’s block.

I sat down over the long weekend to try these all out. And then I remembered all the little tools I use: my guitar tuner battery was dead so I found an online tool. I couldn’t remember where I put my metronome so I found a little app. And I opened my notebook to write down some lyric ideas. I remembered how hard it is to click start on a metronome while holding a guitar, while writing a word, while hitting the record button.

So… yes, had to build a thing.

One Song Studio in action

How it works

🎸 Guitar Tuner – Just like that pedal you have but in your browser

⏱️ Metronome – Set your BPM, time signature, even swing

📝 Word Saver – Capture words as you think of them

🎤 Recorder – High-quality audio recording and playback

🤖 Voice Assistant – Voice-controlled interface using ElevenLabs

Everything works together through natural language commands. Say “set BPM to 120” or “add word catastrophe” or “start recording” and watch it happen.

Notes from Building

Voice + tools is powerful. Last week we started to scratch the surface using ElevenLabs and similar tools to add to cart and filter product. Well, we tried that for a full dashboard: set_metronome_bpm, start_metronome, stop_metronome, set_metronome_time_signature, add_word, start_recording, stop_recording, start_tuner, stop_tuner. Our voice assistant can do it all.

Always interesting to find what the LLM just can’t do. It’s always interesting to stumble across something an LLM is really bad at. Turns out: Tuner – Claude couldn’t do it. Because this isn’t a language problem, it’s a signal processing problem that requires specialized algorithms. Sometimes you just gotta do it yourself (or google a library like aubio.js that does it).  

Short and sweet voice controls. At first the agent was WAY too verbose. We want minimal interruption. The system prompt took a few rewrites to get it to just quickly pop in and out, instead of a long, drawn-out chat while we’re trying to think and record. Hey bot, stop talking!

Word banks help. How to Write One Song is full of exercises to get you unstuck. Many involve writing down favorite words or “stealing” them from books or even overheard conversations. We added a simple word bank so you can capture ideas without putting the guitar down.

Voice-first design = accessibility practice. Designing something to work without hands is a good exercise in accessibility. We’ve seen good hands-free design before (your car is probably pretty good at it) but web tools are still clunky. Ever tried to use a screen reader? Gets tedious fast. And too many sites skip over the WCAG standards that would make them work well. This got our gears turning: what other audio interactions could make the experience smoother and more usable on the web?

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.