Challenges on level calibration of online listening test: a proposed subjective method

In addition to many disruptive consequences in society, the COVID-19 pandemic has also posed challenges on experimental research. The resulting limitations on gatherings of people impeded the attendance or participation of human subjects in experiments. In the context of subjective assessment of sound stimuli by people, listening tests in a laboratory could in principle be replaced by online listening tests, which are moreover more easy to organize for larger amounts of subjects. However, in case of online presentation of sounds, the test environment is not controlled and different apparatuses can introduce a bias in the results. For listening tasks involving sound source localization, compared to loudspeakers, the use of headphones and auralization of sounds taking into account the Head-Related Transfer Function (HRTF) are beneficial. Some psychometric listening tests require a particular excitation level in order to guarantee the consistency of the results over different test people and conditions. When a listening test is offered online, then the listening people typically do not have measurement tools around for reliable quantitative level calibration. The question is whether a subjective calibration method could be developed, which is based on the possible ability of a listening person to equalize a given stimulus to a defined level, based on his or her acoustic memory. In this work, an unsupervised subjective method for level calibration of online presented sound has been investigated on 17 test persons, using pre-recorded speech of a female speaker as a reference signal. The subjectively iterated level was then determined by making use of calibrated reference headphones. The accuracy of the proposed method relies on the classification of the participant practice in terms of speech loudness using a survey prior to the test procedure. The described procedure, which is easy to implement and requires only a few minutes, was found to yield a prediction accuracy of ± 3.8dB.