A handful of German inventors from the Fraunhofer Institute in white coats invent a revolutionary process against all odds to compress music files to one-twelfth of their original size compared to CD with virtually no loss of quality. When was the moment they felt : Are we doing something bigger here?
There are several moments. When I was still a student at the University of Erlangen in 1988 and doing basic research, someone visited our laboratory. My PhD supervisor, Dieter Seitzer, proudly demonstrated to this guest what we were currently working on: compressing digital music files. And when he asked what could become of our work, I replied: “Either our work will be forgotten and it will be accumulating dust in the library, or technology will become a standard that will be used by millions of people.” But I did not dare to dream about it. that really happened.
In 1977, his PhD supervisor, Seitzer, from Erlangen, had the idea of transmitting music by telephone wire. And they all said, “I can’t.” And then you came. What application did you originally have in mind? Was it music in your pocket?
Back then, all textbooks said that you could compress images, videos, and voice, but definitely not music. It is too sensitive and complex. That was the starting point.
We asked ourselves: How can we compress music in that way, that is, reduce the amount of data per piece of music, so that people don’t hear the difference?
The question is to understand how the human ear works so that very similar things happen in our encoder, which compresses the music, as in the inner ear. Even in the inner ear, not all data is transmitted to the brain through nerve fibers. The brain always compares pitches with an internal reference, basically checking what it knows. In addition, there are so-called masking effects: if the sensory hairs tremble in the ear, the other sensory hairs are also automatically stimulated. This leads to the fact that the tones overlap and cannot be perceived at all. This is due to the mechanics of the inner ear. We use this as a guide when we come to the question: For what data can we reduce the level of detail, without being heard? Where would a coarser data structure be acceptable? We did not invent this trick in Erlangen. We weren’t the only ones working on it. We have only brought this knowledge to concrete results faster and optimized it better.
Is it true that you bought records for 1,000 marks in a music store in Erlangen to have compression material?
It is true. We had requested the project and absolutely needed better speakers, a small sound booth, and most of all, lots of audio samples. So I went to buy records: simple pieces, complex pieces, music of all genres, in all areas. We didn’t know what would work and, more importantly, what wouldn’t.
You mean the famous example of the Suzanne Vegas song “Tom’s Diner”, whose a cappella intro with “Da da da da …” was used to fine-tune the psychoacoustic MP3 model. What exactly was it about?
That was a special challenge: dense tones that the ear can still filter very well. My dissertation was almost done at the time and I really believed: I’m done, my process works for all kinds of music. But then I read in a hi-fi magazine that Suzanne Vegas’ voice had been used to test speakers. A colleague bought the CD because we wanted to know: What happens if we compress this music? The result was a disaster.
And how did you solve the problem?
There were two solutions. The first was to realize that what we had read in the specialized literature about how the masking of signals so rich in spectra works was not really true. Then we realized that psychoacoustics in these cases works differently than what the publications of the time suggested. We then test what happens when we transmit the lower frequencies very precisely and become less complex at the higher frequencies in favor of less storage space. That worked
