When you upload a recording to Transcription Studio, a chain of four processing steps kicks off. Each one solves its own problem — and each has its own pitfalls. The result is a sheet of music you can edit, export and share right away.
From signal to spectrum
First, the audio signal is decomposed into its frequency components. The waveform becomes a spectrogram: a map showing which frequencies sound how loud at which moment. For the model this is the foundation — it “sees” music instead of hearing it.
Detecting pitches
A neural network scans the spectrogram for pitches. That sounds simpler than it is: overtones, reverb and several instruments sounding at once overlap into a dense web, from which the model has to extract the notes that were actually played.
The hardest question is not which tones are sounding — but which ones are meant.
Quantizing rhythm
Humans don’t play like machines — and that’s a good thing. To turn freely played timing into readable notation, quantization fits every note into a musical grid:
- Tempo and time signature are estimated from the recording.
- Note lengths snap to sensible values without losing the groove.
- Pickups, triplets and rests get their correct spelling.
The finished score
Finally, the detected notes become a cleanly engraved sheet — with clefs, accidentals and instrument-specific notation. You can fine-tune the result in the editor and export it as PDF, MIDI or MusicXML. Or just try it yourself:
Read more about the research behind the transcription pipeline on our research page.