Combining Probabilistic and Unsupervised Transcription of Piano Music
Unsupervised Transcription of Piano Music
The ability to translate human musical expressions into notational systems has long been an important goal in the field of music transcription. This process, however, remains a complex and difficult task. This is especially true for piano music where the acoustic signal can be very complex due to the many possible playing techniques and material characteristics of the instrument. The distinct timbres of different instruments are also a challenge for automated transcription systems as they can introduce significant variation into the frequency range of the audio signal.
One common approach to piano transcription involves using unsupervised learning based on sound features. The acoustic features are typically computed from spectrograms obtained from a sound recording of a pianist playing a musical passage. The resulting features are used to train a model that performs note level transcription of the piano music. The model is evaluated with respect to a standard benchmark.
Another method for piano transcription is to use a probabilistic approach based on the probability distribution of note events within a given time frame. This approach can be used to construct a score of notes that are arranged in the correct time order. The score can then be converted to music notation that can be directly input into a computer keyboard.
Combining Probabilistic and Unsupervised Transcription of Piano Music
The problem with these methods is that they do not address the complexities of note-level transcription such as vibrato, articulation, and other musical nuances. Moreover, these methods tend to be less accurate in the presence of non-musical acoustic variations such as noise and reverberation.
A number of researchers have been attempting to improve transcription systems by combining both the probabilistic and unsupervised approaches. A recent paper presents a model that utilizes both methods to achieve better results than existing state-of-the-art systems. The authors show that their model is capable of predicting both the pitch contour and the note-level transcription of piano music in noisy and reverberant conditions.
Some classical musicians have been reluctant to take the Liszt transcriptions of Beethoven symphonies seriously. But Glenn Gould, in a remarkable recording of the first movement of the Fifth Symphony, shows that the piano can make these pieces seem fresh again by offering textures, inner voices and grand structure without the orchestral colorings that can obscure them.
More recently, adventurous younger pianists have been exploring transcriptions of music from other historical eras. Jeremy Denk’s recent album for Nonesuch featured his piano versions of medieval and Renaissance vocal pieces by Machaut, Ockeghem and Josquin. Behzod Abduraimov began his 92nd Street Y recital with Liszt’s transcriptions of Wagner’s “Liebestod” and the “Solemn March to the Holy Grail.”
The key to success with any type of piano transcription is to work in small chunks. Trying to transcribe an entire musical piece at once will overwhelm the listener and lead to mistakes. Instead, focus on a single bar or two bars of music and attempt to capture what is being played in as much detail as possible. This approach will allow you to transcribe more accurately and faster while also freeing up your brain to focus on other things such as playing the piano.