The development of MPEG-1 - Part B

 Last update: 2005/05/01

Work in the Audio group was also progressing. A sizeable portion of the group was made up of people who were interested in audio-only applications, some of them working in the Eureka 147 DAB project. For them it was important to develop a standard that would provide CD-quality at the bitrate of 256 kbit/s. This target affected the video work because video simulation results had to be shown at 1.15 Mbit/s because this was the remaining bitrate from the total payload of 1.4 Mbit/s of CD. 

The approach of the Audio group in the development of the standard showed some differences from the one used by the Video group. Instead of producing a general CfP, the Audio group first worked to cluster the proposals that the different companies were considering (of course the CfP was to be open to anybody else). 

These were the four clusters: 

1. Transform Coding with overlapping blocks 
2. Transform Coding with non-overlapping blocks 
3. Subband Coding with less than or equal to 8 subbands 
4. Subband Coding with more than 8 subbands. 

The clusters were encouraged to provide a single proposal and this indeed happened. Swedish Radio (SR) was kind enough to perform the subjective test of the four clustered proposals using "golden ears", the name given to specialists capable of detecting the slightest imperfection in sound. The results of the subjective tests were shown in Stockholm in June 1990 (this was formally a part of the Porto meeting, where the rest of MPEG was meeting). The reason for having this meeting in Stockholm was to be able to listen to the submissions in the same setup the golden ears had used for the tests. 

The first clustered proposal performed the best in terms of subjective quality but, not unexpectedly, the implementation penalty was higher than the fourth clustered proposal that scored less but with a lower implementation complexity: an undoubted challenge, which the audio chairman resolved with time and patience. This was the last achievement of Hans Mussmann who left MPEG at the Paris meeting in May 1991. His place was taken over by Prof. Peter Noll of the University of Berlin. 

The result of the work was an audio coding standard that, unlike the corresponding video standard, was not monolithic because there were three different "flavours": the first - called Layer I - was based on subband coding and had low complexity but the lowest performance, the second - called Layer II - was again based on subband coding with average complexity and good performance and the third - called Layer III - was based on transform coding and provided the best performance, but at a considerable implementation cost. So much so that, at that time, many considered Layer III as impractical for a mass-market product. There could be 3 different conforming implementations of the standard, one for each layer. The condition was imposed, however, that an MPEG-1 Audio decoder of a higher layer had to be capable of decoding all the lower layers. 

The verification tests carried out before releasing the MPEG-1 Audio standard showed that subjective transparency, defined as a rating of the encoded stereo signal, as assessed by "golden ears", greater than 4.6 in the 5-point CCIR quality scale, was achieved at 384 kbit/s for Layer I, 256 kbit/s for Layer II and 192 kbit/s for Layer III. The promise to achieve "CD quality" at 256 kbit/s with compressed audio had been met and surpassed. Today, with continuous improvements in encoding (which is not part of the standard) even better results can be achieved.

The development of the Systems part of the standard was done using yet another methodology. After determining the requirements the Systems layer had to satify, the group, a most diversified collection of engineers from multiple industries, decided that they did not need a CfP, because the requirements were so specific that they felt they could simply design the standard by themselves in a collaborative fashion. The initial impetus was provided by Juan Piñeda, then with Apple Computer, at the Porto meeting in July 1990, when the first packet-based multiplexer was proposed by him. Eventually Sandy MacInnis of IBM became the chairman of that group after Allen Simon's resignation. 

One of the issues the group had to deal with was "byte alignment", a typical requirement from the computer world that the telco world, because of its "serial" approach to bitstreams, did not value, as shown by H.261 whose bitstreams are not byte aligned. Byte alignment was eventually supported in MPEG-1 Systems because the system decoding of a byte-aligned 150 kbyte/s stream was already feasible using the CPUs of that time. In the process, the MPEG-1 Video syntax, too, was made byte aligned. 

Another issue was the choice between constant and variable packet size. One could have thought that, because the main target of MPEG-1 was DSM where disk formats have a fixed block size, a fixed packet length should have been selected. Eventually, however, a variable packet size was selected, a consequence of the fact that the physical format of the disc is, in OSI terminology, a layer 2 issue, while packet multiplexing is a higher-layer issue that did not necessarily have to relate with the former. In conclusion MPEG-1 Systems turned out to be a very robust and flexible specification capable of supporting the transfer of tightly synchronised video and audio streams across an arbitrary error-free delivery system. 

One morning of July 1990, Arian Koster of KPN Research called me to make a suggestion: "what if MPEG developed a software implementation of the MPEG-1 standard?" I immediately asked what would be the benefits for MPEG if we had done that. He said that various companies had already developed their own software implementations of the Video SM, because that was necessary to take part in CEs. More software would also be developed for Audio and Systems. If everybody would give just a small portion of their code, MPEG could have had a complete software implementation of the standard and everybody in MPEG would have gained from being able to access it. Frankly, I did not see at that time for what reasons anybody should have given away part of his code, but it has always been my policy not to disallow something other people believed in just because I did not understand it. I never had to regret this policy and certainly not in this case, which created the seeds for one of the major innovations in MPEG, as we will see later. 

Slowly, the idea worked and already at the first Santa Clara, CA meeting in September 1990 the Audio group, made of the most contentious people in MPEG, but open to novelties and structured in their implementations, had already proposed an ad hoc group on "Software Simulation - Audio". With the contribution of many, in 1994 MPEG could release part 5 of MPEG-1: "Software simulation" (part 4 had already been assigned to "Conformance Testing"). While the first four parts of MPEG-1 are normative, in the sense that if you want to make conforming products or bitstreams they must satisfy the requirements of the relevant parts of the standard, part 5 is a Technical Report (TR), i.e. something that is produced for the general benefit of users of the standard, but has no normative value. It is, in ISO language, "informative". 

The second MPEG London meeting in November 1992 put the seal on MPEG-1, with the approval of the first three parts of the standard: Systems, Video and Audio. Since then the standard has been very stable: in spite of its complexity, very few corrigenda were published after that date: two for Systems, three for Video and one for Audio. The MPEG-1 work did not stop at this meeting, however, because work on conformance and reference software continued well into 1994. 

I would like to conclude this snapshot on MPEG-1 by highlighting the role that my group at CSELT had in developing the first implementation of a full MPEG-1 decoder. Already in 1984 we had started designing a multiprocessor architecture for video coding based on a bus interconnecting a set of multiprocessor boards. This had first been used in 1986 to implement one of the first ISDN videophones that was made a product by an Italian company and actually used in trial services. Each board featured four Analog Device 2900 Digital Signal Processors (DSP) and one Intel 80186 that controlled communication between the DSPs and between the boards, because each board had in charge only a slice of the picture and data had to be passed to the different boards by virtue of Motion Compensation. 

As part of the CSELT work in the COMIS project, my group extended this architecture and implemented the first MPEG-1 Systems, Video and Audio decoding in real time (the MPEG-1 Audio decoding board had been purchased from Telettra) and demonstrated it at the Haifa meeting in March 1992 (figure below). 

The COMIS demo

This was not just a technology demonstration because other partners in the COMIS project (BBC, CCETT and Olivetti) had developed content that included multimedia navigation support based on the MHEG standard, then still under development (an example can be seen on the screen in the figure above). 

I would like to close this chapter reaffirming that the MPEG-1 standard has been created by the subgroup chairs during interminable sessions where thousands of technical contributions were debated, by some natural leaders who sprang up and took the lead to resolve thorny issues and by the hundreds of people who took part in all these discussions. I created the environment for this to happen, they have made the standard.

Copyright © 2003 chiariglione.org