Acoustic Guitar as Percussion
A Mel-Spectrogram Blind Spot
On a mel-spectrogram, a drum hit and a guitar strum can look remarkably similar, and pretrained audio taggers such as PANN often confuse the two, classifying acoustic guitar attacks as percussion.
This is not simply a labelling problem or a matter of insufficient fine-tuning. The limitation begins in the base mel pipeline. The STFT introduces a time-frequency trade-off, while the mel filterbank provides only coarse resolution at higher frequencies. These are followed by additional lossy operations such as logarithmic compression. Together, these processing stages discard the transient structure and high-frequency detail that distinguish a percussive hit from a strummed attack. Because this information is removed before the network ever processes the signal, it cannot be recovered through downstream fine-tuning.
Since mel-based representations form a standard starting point for most modern audio tagging systems, this limitation is inherited across a large body of pretrained models rather than being confined to a single architecture.
This talk examines why this confusion occurs and surveys alternative audio representations designed to preserve transient information, including source separation as preprocessing, adaptive compression methods such as PCEN, CQT, and learned neural encodings. Rather than presenting a single solution, it explores the trade-offs each approach makes between computational cost, robustness, and classification performance.
Shreya Gupta
Shreya Gupta is a musician, currently pursuing a Master’s in Sound Design at the University of Edinburgh. With a background in Indian classical music, composition, and audio programming, her curiosity has led her to experiment at the intersection of tradition, technology, and cross-cultural collaboration.