AccompaniX 2017: Artificial Duet Performer Competition

AccompaniX is a competition in computer-generated accompaniment for human-computer duet performance. All duet pieces contain two parts: a human lead (to be performed by a human) and a computer accompaniment (to be performed by the submitted artificial performer). We mechanize the scenario in which the two performers, who are allowed to practice for a while, arrive onstage for a new performance. Though both performers “know” the music they will play, their performance will differ from others in that the accompanist needs to follow the leader anew, reacting to the human performance in a musical manner. In particular, the human lead part is an expressive rendition of a pre-defined score, the computer accompaniment has the freedom to follow a pre-defined score or to “compose” its own accompaniment score.


  1. Participants will be given 4 duet pieces (see an example here) on May 29, 2017, in the form of (1) a score (2) a human lead performance of the duet. The latter will be in the form of MIDI and audio files, or the participants are free to create their own human performance.
  2. Participants shall choose (at least) one of the 4 pieces. The “test” is then when given a new performance version of the human lead part, the machine (artificial performer) shall perform the accompaniment musically in concert with the human lead part. Computer accompaniment has the freedom to interpret the pre-defined score or to “compose” and perform its own accompaniment part.
  3. After the 4 duet pieces are released, participants will have 72 hours to construct their artificial performers (or to “train” the artificial performers if any machine-learning techniques are involved) and to then produce a performance. The performance could be in the format of either MIDI or audio.
  4. The created artificial performer should be a Linux/Windows/Mac executable program (written in C, Python, etc.) capable of sequentially processing the human part and generating a computer part. Note the program should be causal; during the course of a performance, artificial performers cannot use future human performance as a guide to generate their current performance.
  5. Validation will take the form of giving the artificial new human performances (of the same piece) to follow in the form of MIDI or audio input.


  1. Contest participants can choose to submit entries that derive from any one of the 4 pieces.
  2. Submissions comprise (1) the source code for the artificial performer, (2) the final mix-down which contains both the human lead part and the generated computer accompaniment part, and (3) the human lead performance alone, if the participants choose to create their own human performance rather than use the given one.
  3. Submissions will be effected through access to a virtual machine running a Linux/Windows operating system, or a server running Mac OS X. Participants will have access to the machines days before the competition begins and are responsible for getting their software to run on the platform (although the organizers will do their best to help if problems arise). See system specifications here for available operating systems and timelines to access them.
  4. The actual data for competition will be released on May 29, 8 am EST. Entries must be submitted by 8:00 AM EST, June 1, 2017.
  5. The virtual machine should be able to run the code via a single command line (or an equally simple and directly operation), which takes human lead performance as the input and generate computer accompaniment output, using less than 16GB of RAM and less than 50GB of disk space.
  6. The program must be completely self-contained and require no connectivity to the Internet.

Get started with your application here.

Prize and Honor:

  1. A set of finalists will be determined .Among the finalists, the best program (as determined by a panel of judges through Turing Test) will receive a $2000 cash prize. Judges will give prizes for outstanding creative efforts at their discretion.