GPT-6 Astra created an SVG spectrogram of "Daisy Bell" that can be played as audio.
Max Rubin had tested whether GPT-6 could identify sounds by looking at a spectrogram as an image. It could. So I took the next step. If it can hear spectrograms, maybe it can speak with them as well.
I first asked it to draw a spectrogram of "Mary Had a Little Lamb." It worked, but after auditing the logs, I found that it had written a script defining the frequencies for different phonemes. That wasn't quite what I wanted.
I started a fresh thread and made the rules clearer. It had to write the SVG manually. No script generating the phonemes, no copying an existing spectrogram, and no simplifying the task to just the notes without the words.
It could use code to convert its SVG to audio and check the result with transcription, then revise the SVG manually.
The result is rough. Transcription missed "Daisy" and several lines. But it produced partly intelligible singing by writing an SVG.
You can see the prompt and download the SVG, along with the earlier attempt.
We haven't scratched the surface of this model's capabilities.