Speaker
Description
Identification of musical instruments has primarily been investigated for sounds in isolation. However, as musical scene analysis often involves multi-source scenarios, research must be extended to sound mixtures.In this study, thirty-four listeners identified sounds from eight Western classical instruments presented in isolation and in stationary chord mixtures of two to four instruments. For the chord mixtures, the task required listeners to hear out and identify a target instrument within the mixture, guided by a preceding sinusoidal cue. Behavioral results showed that identification accuracy decreased as the number of instruments in the mixture increased. Furthermore, results revealed instrument-dependent confusions and effects of register and voice positioning within the chord.To further explore these findings, we implemented a computational frontend-backend model. In the frontend, a harmonic mask containing the first ten harmonics of the target F0 was applied to the mixture spectrograms to extract target instrument representations. The width of these harmonic windows served as a parameter for frequency resolution in pitch tracking. Following auditory ERB-filtering, spectral and temporal modulations were captured using the modulation power spectrum (MPS). For the backend, a random forest classifier was trained on isolated instrument sounds and tested on the masked mixtures. Internal noise was introduced as a second parameter to impair pattern matching.Our model provides plausible classification results that reflect human performance in both accuracy and confusion patterns of instruments, establishing a foundation for understanding the auditory perception of instrument timbre within complex musical mixtures.