Determine the cause of the MAESTRO audio confound

Determine whether clipping during the loudness adjustment caused the statistical distinction between attended and ignored recordings in the MAESTRO candidate speech signals, or identify the actual cause of that audio mark, using a dataset or experimental design that separates recording identity from attended and ignored roles.

Background

The paper finds that scale-invariant properties of the candidate speech recordings—such as peak-to-average ratio, skew, sparsity, and frequency-band balance—can reveal which talker was attended, even after loudness standardization. The dataset authors attribute this distinction to clipping of the attended talker’s peaks during loudness adjustment, but the present paper does not independently establish that explanation.

Because each recording is used in only one role, the released MAESTRO data cannot disentangle whether the observed mark is caused by the recording itself or by its attended-versus-ignored role. Resolving this issue would clarify the source of the benchmark confound and motivate a cleaner dataset design in which each recording appears in both roles.

References

The dataset's authors explain the mark by clipping during the loudness adjustment \citep{maestro2026}, but the released data cannot test this directly.

— Merit: Making every modality earn its contribution in multimodal attention decoding  (2610.17507 - Alavi et al., 13 Oct 2026) in Appendix, Section “Further limitations,” paragraph “The cause of the audio mark is not tested directly”