Models for evaluating navigational techniques for higher-order ambisonics
Type
Models are presented that predict perceived source localization and spectral coloration for the purpose of evaluating navigational techniques for higher-order ambisonics. Previous evaluations have typically relied on binaural localization models, which conflate the effects of the navigational technique with those of the adopted ambisonics-to-binaural rendering approach. Moreover, studies on navigation-induced coloration have been largely qualitative. The presented models are applied directly to translated ambisonics impulse responses (i.e., before rendering to binaural) and are validated through listening experiments. Localization is predicted using an extension to a precedence-effect-based localization model. Coloration is predicted using a linear combination of spectral energies and notch-depths in a difference-spectrum between the test and reference signals. For two interpolation-based navigational techniques and a range of translation distances, localization and coloration are also measured subjectively through binaural-synthesis-based listening tests, wherein subjects judge source position for a spatialized sample of speech and rate the induced coloration in pink noise relative to reference signals. The proposed localization model is shown to predict the data with comparable accuracy to that of a binaural localization model and the coloration metrics used are shown to best predict perceived coloration compared to alternative sets of metrics.
Errata:
Model Name |
Residuals |
Correlation |
¯ϵϵ¯ |
---|---|---|---|
Proposed |
706.5 |
0.85 |
3.67° |
Dietz et al. (2011) |
1009.2 |
0.82 |
4.34° |
- On page 10, the Pearson correlation coefficient for the proposed model given in Table 3 should read 0.85. Consequently, Table 3 should read:
- Accordingly, the following discussion (on page 10, below Eq. (19)) should read:
"From these values, we see that the proposed model seems to fit the data better compared to the binaural model, as the former achieves a lower squared-residual value, a higher correlation with the data, and a smaller mean absolute prediction error. This is surprising given the binaural model's ability to take into account subject-dependent variations, since the predictions are made on a per-subject basis (see Sec. 2.A.ii), as well as any effects of the binaural rendering approach used."
Copyright notice: Copyright (2017) Acoustical Society of America. This article may be downloaded for personal use only. Any other use requires prior permission of the author and the Acoustical Society of America. The following article appeared in Proc. Mtgs. Acoust. 30, 050009 (2017) and may be found at https://doi.org/10.1121/2.0000625.