Neural networks trained on one subject’s data can predict human sound-elevation localization

Evidence

As seen in Section 12.3.4, the localization of a sound source in the median plane relies on spectral cues. Generically, a machine-learning approach can be used to simulate human performance in localizing the elevation of a sound source. In particular, an accurate prediction can be achieved using artificial neural networks trained with large data sets from a single subject (Jin et al., 2000). Recently, a functional model that predicts listeners’ performance in such localization tasks has been proposed by Baumgartner et al. (2013). This model computes the probability of perceiving a sound coming from a certain elevation angle by comparing the internal representation of the sound as processed by the peripheral auditory system with an internal representation of the HRTF. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 1687

Categorization

Category: Factual Claims

• XREF: Machine-learning simulation of perception connects to the reader’s broader interest in acoustic communication applications; HRTF spectral cues are the mechanism behind 3D audio rendering.

Related

• Neural networks trained on single-subject data can accurately model human elevation localization (similarity 0.98)
• Spectral elevation cues predict human sound-localization performance via machine learning (similarity 0.90)
• Sound localization by elevation is guided by plausibility, not just acoustics (similarity 0.81)
• Elevation localization relies on pinnae spectral cues encoded as HRTFs (similarity 0.80)
• Large diverse datasets of human HRTF and ear geometry are now being collected (similarity 0.80)
• Horizontal localization uses binaural cues while vertical localization does not (similarity 0.79)
• Elevation localization is a timbral recognition task driven by resonant peaks (similarity 0.79)
• Humans localize sound accurately using only two real-time ear signals (similarity 0.78)