A new version of our scanning model now reads every PDF and photo import. It is built on a new architecture that looks at each line of music in more detail, and it was trained on a cleaner set of examples. It is more accurate on 24 of the 26 kinds of music we test it on.

• Ties and slurs are better told apart. A curve joining two different notes is now much less often read as a tie that changes the second note's pitch.
• Silent staves are read more reliably. In a full score, a staff that rests for a whole line is less often filled with invented drum notes or dropped from its part.
• Long rests are counted more closely. A run of 20 or more empty bars comes much closer to its printed length, instead of turning into dozens of extra bars.
• Fewer wrong notes. About 8% fewer pitch mistakes, 17% fewer tuplet mistakes and 10% fewer chord mistakes than the previous version.
• Notes are located more precisely on the page. This helps lyrics, dynamics and other markings attach to the note they belong to.