In 1987, the Fraunhofer IIS started to work on perceptual audio coding in the framework of the EUREKA project EU147, Digital Audio Broadcasting (DAB).
In a joint cooperation with the University of Erlangen (Prof. Dieter Seitzer), the Fraunhofer IIS finally devised a very powerful algorithm that is standardized as ISO-MPEG Audio Layer-3 (IS 11172-3 and IS 13818-3).
MPEG-3 =Moving Picture Experts Group Audio Layer – 3
The 3 MPEG layers Three Layers, Three Applications Name Compression Factor Bit rate Application Layer 1 1:4 384 Kbit/sec. Digital Compact Cassettes Layer 2 1:6 to 1:8 256 to192 Kbit/ sec. Digital Radio Layer 3 1:10 to 1:12 128 to 112 Kbit/ sec. Digital Internet Music
MP3 Performance An Overview of MP3 Quality Levels Sound Quality Mode Bit rate Compressions Rate Telephone mono 8 Kbit/s 96:1 Better than SW Radio mono 16 Kbit/s 48:1 Better than MW Radio mono 32 Kbit/s 24:1 Similar to VHF Radio stereo 56 to 64 Kbit/s 26 to 24:1 Similar to CD stereo 96 Kbit/s 16:1 CD quality stereo 112 to 128 Kbit/s 14 to 12:1
The audio signal passes through a filter bank which divides the audio signal in 576 areas (sub-bands). This requires very complex filters. Here, MP3 encoders work with the well-known Discrete Cosine Transformation.
In real time it calculates unnecessary frequencies and eliminates them iteratively (repeatedly) until the best possible result is achieved.
At the same time, the audio signal passes through the psychoacoustic model. For every sub-band of the entire signal spectrum, the masking threshold is determined using the Discrete Fourier Transformation.
Joint stereo coding can then be done to take exploit the fact that both channels of a stereo channel pair contain by far the same information. These stereophonic irrelevancies and redundancies are exploited to reduce the total bitrate.
When quantizing the sample another starting point for data reduction arises. Every sample is made up of 16 bits, but not all 16 are necessarily needed in order to represent the sound. As such, the leading nulls of a 16-bit sample may be left out
At the same time, individual samples are analyzed and compressed again using Huffman encoding. This produces a further reduction of data of about 20 percent.