RECOGNITION OF ISOLATED INSTRUMENT TONES
BY CONSERVATORY STUDENTS

Asha Srinivasan, David Sullivan, and Ichiro Fujinaga
Peabody Conservatory of Music
Johns Hopkins University

Background
Musicians have a remarkable ability to recognize instruments by timbre;
however, previous experiments using isolated tones suggest that
recognition rates range between 36.5% and 90.0%. Recently,
timbre-recognition computer models have been able to match or exceed these
rates.

Aims
Using conservatory students, this experiment aims to reconstruct previous
experiments, to measure the effect of ensemble experience and short-term
training on the recognition rate, and to generate more detailed baseline
data to help evaluate computer performance.

Method
Two tests were performed. The first test included four sections, involving
2, 3, 9, and 27 instruments. The instruments involved in each section were
given on the answer sheet, and then isolated tones played by those
instruments were presented in random order. The subjectsÕ task was to
identify which instrument had produced each tone. The second test repeated
the first test with only three parts, involving 2, 9, and 27 instruments.
Additionally, short training sessions preceded each section (lasting 2
min., 3 min., and 10 min., respectively). In these training sessions, each
instrument was identified before sounding, several tones from each
instrument were played, and brief tests were interspersed to maintain the
subjectsÕ engagement. Eighty-eight subjects participated in the
experiment. All tones were taken from the McGill University Master
Samples.

Results
In our tests, average scores were: 94.5%, 97.6%, 90.2%, and 55.7% for 2,
3, 9, and 27 instruments, respectively. Additionally, subjects who play
orchestral instruments scored significantly higher than those who do not.
Finally, the short training sessions had no significant effect on the
subjectsÕ performance.

Conclusions
Compared to previous experiments, the average scores of subjects in this
experiment were considerably higher. Furthermore, subjects who play
orchestral instruments tended to score higher than those who do not. The
excellent average score of the human subjects in this experiment presents
new challenges for timbre-recognition computer models.

Topic Areas
Timbre recognition, musical training, computer recognition, computer 
models