MOST of us know better than to accept advice from an ignoramus. Yet we are content to listen to music playlists or even to buy tracks based on a computer鈥檚 assessment of what we might enjoy, despite the software not knowing the first thing about music. The plan now is to make this less of a hit-and-miss affair by giving computers a course in musical appreciation.
Online music stores base their recommendations on similarities in buying or browsing patterns: when you buy a song, for example, the system might highlight music purchased by others who have bought the same song as you. Other online music services, such as Pandora, combine this so-called crowd-filtering approach with human labelling 鈥� where users assign descriptive terms to a song. The system can then choose songs for you whose descriptions match those of a track you have already selected, for example.
These services work up to a point, but they tend to favour artists who already have significant sales data to work with, or else they rely purely on the opinions of human labellers, says , an artificial intelligence researcher at the University of California, San Diego. Could we improve such services by teaching computers the rudiments of music?
Advertisement
Barrington is building software that can analyse a piece of music and distil information about it that may be useful for software trying to compile a playlist. With this information, the software can assign the music a genre or even give it descriptions which may appear more subjective, such as whether or not a track is 鈥渇unky鈥�, he says.
鈥淭he software can give the music a more subjective description, such as whether or not it is funky鈥�
Before any software can recommend music in this way, it needs to be capable of understanding what distinguishes one genre of music from another. Early approaches to this problem used tricks employed in speech recognition technology. One of these is the so-called mel-frequency cepstral coefficients (MFCC) approach, which breaks down audio into short chunks, then uses an algorithm known as a fast Fourier transform to represent each chunk as a sum of sine waves of different frequency and amplitude. In speech recognition, the nature of the sum arrived at can help a computer distinguish between phonemes, the units of sound that make up speech. When applied to music, the technique can be used to assign loose categories to songs.
The technique is useful for determining which instruments are being used in a piece but is otherwise pretty basic, says , who researches computer audio analysis at Columbia University in New York. 鈥淚t鈥檚 like taking a 10-megapixel image and representing it as a blurry thumbnail,鈥� he says.
and colleagues at the University of S茫o Paolo in Brazil take a different approach. Their software looks at rhythm to assign a genre to music.
Unlike melody, rhythm is potentially a useful way for computers to find a song鈥檚 genre, da F. Costa says, because it is simple to extract and is independent of instruments or vocals. Previous efforts to analyse rhythm tended to focus on the duration of notes, such as quarter or eighth-notes (crotchets or quavers), and would look for groups and patterns that were characteristic of a given style. Da F. Costa reasoned that musical style might be better pinpointed by focusing on the probability of pairs of notes of given durations occurring together. For example, one style of music might favour a quarter note being followed by another quarter note, while another genre would favour a quarter note being succeeded by an eighth note.
By analysing a collection of MIDI files 鈥� electronic transcriptions of music 鈥� da F. Costa鈥檚 team was able to establish models of the note transitions characteristic of rock, blues, reggae and bossa nova songs. They then tested these models against a selection of tracks available on last.fm that users had already suggested a genre for. For tracks that had been placed in a single genre by users, the team鈥檚 results agreed with the user-suggested genre in 71 per cent of cases (New Journal of Physics, ). While not perfect, the result does suggest that rhythm can play a useful role in categorising songs, says da F. Costa.
Barrington, however, believes that assigning genres to entire tracks suffers from what he calls the Bohemian Rhapsody problem, after the 1975 song by Queen which progresses from mellow piano introduction to blistering guitar solo to cod operetta. 鈥淔or some songs it just doesn鈥檛 make sense to say 鈥榯his is a rock song鈥� or 鈥榯his is a pop song鈥�,鈥� he says.
Barrington wanted to create a system that could distinguish between different styles of music within a single song. He started out using the MFCC approach on songs that did not vary in style, breaking these songs down into overlapping 5-second chunks and constructing a profile of how the MFCC features evolved in the course of each song. This was used to create a genre 鈥渇ingerprint鈥� for whole songs by matching the profile to those in songs volunteers had labelled. Last year he showed that this system was capable of taking a 鈥渟eed鈥� song and producing a related playlist that users rated as highly as one generated from buying patterns.
Now Barrington has applied that approach to identify styles within a single song. If a user has chosen a song with a mellow verse and a raucous chorus, for example, the system can recommend songs that follow a similar pattern, rather than merely being in the same genre. The work will be presented at the International Society for Music Information Retrieval Conference in Utrecht, the Netherlands, this August.
While such efforts show that computerised music analysis is making progress, it also highlights how unsophisticated these systems are today, says Ellis. 鈥淲hat they鈥檙e doing is much less than what any reasonably musically literate human can get out of the sound,鈥� he says. 鈥淲hat鈥檚 amazing, though, is that we can build these incredibly crude systems and still do these tasks.鈥�
