The Record You Cannot Actually Hear

A producer in a bedroom in Atlanta drags an mp3 of "What's Going On" into a folder, clicks a button, and four seconds later he has the bass alone. James Jamerson, no Marvin Gaye, no strings, no piano, no horn section, no room. Just the line. He drops it into his DAW, slows it down by half, and starts to learn it. What he is hearing is, in a very real sense, not what Jamerson played.

It is close. Closer than anyone could have dreamed in 2010. The stem splitter on his laptop is using a neural network trained on hundreds of thousands of multitrack recordings, and it has learned the statistical signature of a Fender Precision Bass with flatwound strings. It can pull the bass out of a 1971 Motown mix with shocking accuracy. But it is also doing something else, quietly, beneath the result: it is reconstructing the parts of the bass that were buried under the kick drum, behind the piano, ducking inside the snare. Where the bass disappeared in the original, the algorithm has to make it up. It hallucinates Jamerson. It is good at it. It is not the same as hearing Jamerson.

For thirty years, learning from records meant trying to reverse-engineer something you could not take apart. You sat at a stereo and lifted the needle until you could hum the line, then you crawled inside it with your instrument. Some bass players spent six months on "What's Going On." Some of them came out the other side with the line. Some of them came out with something better — their own version of what they thought Jamerson was doing, which became their voice. The friction of not being able to isolate the part was, it turns out, the teacher.

The friction is gone. The reward is a working bass player at twenty-three who has internalized a thousand basslines instead of fifty. The cost is that the basslines in his ear are processed. They have a slight smear at the attack where the algorithm's phase reconstruction softened the transient. They have a strange airy quality in the upper midrange where the model filled in the harmonics it could not separate. He is not playing along with Jamerson. He is playing along with a remarkable forgery that the world has agreed to call Jamerson.

None of this is a reason to put the tool down. Stem separation is a real gift, and the producers who refuse to use it out of purism are kidding themselves about what learning has always been. The point is to know what the tool does. The young producer who pulls a record apart to study it should also, once in a while, sit with the mix as it was actually pressed — drums and bass and voice and air, refusing to come apart. He should learn the sound of a part doing its job inside a song, not the sound of a part rescued from the song. Because what made Jamerson Jamerson was not the line in isolation. It was the line as it ducked, as it answered, as it left holes for the snare and slid into the gaps between Marvin's phrases. The algorithm cannot give you that. Only the record can.

There is a record on every laptop now that has never existed in the air. Listen, sometimes, to the one that has.