Slashdot Mirror
Researchers Create an Automatic Backup Band for Singers
Researchers at Microsoft Labs are hoping to allow untrained singers to have their own automatic backup band in the near future. A new piece of software, "MySong", promises to take a sung melody and using a probability computation algorithm, generate an appropriate chord accompaniment. There is also a video of the process on the Microsoft Labs website. "'The idea is to let a creative but musically untrained individual get a taste of song writing and music creation,' Morris told New Scientist. 'There was nothing out there that could take a sung vocal melody as an input and then generate appropriate chords to accompany it. [...] Since people rarely sing at precise frequencies, MySong compares a sung melody to the 12 standard musical notes. It then feeds an approximate sequence of notes to the system's chord probability computation algorithm. This algorithm has been trained, through analysis of 300 rock, pop, country and jazz songs, to recognize fragments of melody and chords that work well together, as well as chords that complement each another.'"
...considering how unimaginitive most bands are today - the 1-4-5-1 progression is so prevalent in pop music, you can hum most songs on the radio within the first two minutes of listening to it.
Experiment: pick three Linkin Park songs (from their frist couple of albums), play the first, and sing the melody from the second or third over it. You'll be amazed at how different they aren't.
Quiz: True or False -- On a scale of 1 to 10, what is your middle name?
You're absolutely correct on all points. The system will create "best guess" chord progressions, in any case. I assume it's able to create several guesses based on a probability threshold the user sets initially (probably something like "show me the top 5 most 'fitting' progressions").
Based on what it's trained on, the system will show certain tendencies. If after training it's boxed up and given to a user to work with (no further training possibly by user), then the user will have to learn what these tendencies are and adjust accordingly.
And yes, to not create total rubbish the singer will have to have some musical sense. Just like how a "language model" is used to pick out the most-likely-to-be-correct translation from a lattice that the translation model generates in statistical natural language translation systems, the singer might need to pick out what he/she desires out of a set of possibilities the music generation system presents.
So, if your point was that this system will not be able to instantly fulfill an amateur singer's desires, then you're definitely right. Ideally the system would be able to be further trained on music the amateur singer personally enjoys (or wants to emulate), and would also learn from the choices the singer makes when selecting progressions generated by the system. Over time, then, it would do a better job of mapping the singer's vocals to what he/she wants to hear as an accompaniment.
I like basketball!!1!
That is and it isn't true. As someone who's often responsible for improvising harmonies to match a melody, it's quite possible to provide an interesting harmony to a boring melody and vice versa.
My personal favorite example here is the popular song "Turkey in the Straw". The traditional harmony goes something like this (assuming the key of C):
verse: C-C-C-C-C-C-G-G-C-C-C-C-C-C-G-C
chorus: C-C-C-C-F-F-F-F-C-C-C-C-C-C-G-C
However, this is a very nice more complex harmony:
verse: C-C-C-C-C-Am-Dm-G-C-G/D-C/E-C/E-F-D/F#-G-C
chorus: C-C-C-C-F-F/E-Dm-G-C-D-D#dim-C/E-F-D/F#-G-C
The melody works either way, but the harmonies are quite different.
I am officially gone from
If somebody asks to be accompanied on an unknown song, most musicians will initially try the 4 chord progression known as Rhythm changes (named for Gershwin's "I've got Rhythm"). Often it works, and in listening to MS kludge it seems they likes their Rhythm changes.