Cross-Cultural Bias in Mel-Scale Representations: Evidence and Alternatives from Speech and Music
Abstract
Modern audio systems universally employ mel-scale representations derived from 1940s Western psychoacoustic studies, potentially encoding cultural biases that create systematic performance disparities. We present a comprehensive evaluation of cross-cultural bias in audio front-ends, comparing mel-scale features with learnable alternatives (LEAF, SincNet) and psychoacoustic variants (ERB, Bark, CQT) across speech recognition (11 languages), music analysis (6 collections), and European acoustic scene classification (10 European cities). Our controlled experiments isolate front-end contributions while holding architecture and training protocols minimal and constant. Results demonstrate that mel-scale features yield 31.2% WER for tonal languages compared to 18.7% for non-tonal languages (12.5% gap), and show 15.7% F1 degradation between Western and non-Western music. Alternative representations significantly reduce these disparities: LEAF reduces the speech gap by 34% through adaptive frequency allocation, CQT achieves 52% reduction in music performance gaps, and ERB-scale filtering cuts disparities by 31% with only 1% computational overhead. We also release FairAudioBench, enabling cross-cultural evaluation, and demonstrate that adaptive frequency decomposition offers practical paths toward equitable audio processing. These findings reveal how foundational signal processing choices propagate bias, providing crucial guidance for developing inclusive audio systems.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the identification and quantification of cross-cultural bias in mel-scale audio representations, alongside the introduction of alternative representations that significantly reduce performance disparities. This work is a critical step towards developing fairer audio systems, highlighting the importance of cultural considerations in machine learning applications.
Methodology
The paper employs a robust methodology that systematically evaluates the impact of mel-scale representations on audio processing across diverse cultural contexts. The authors isolate the contributions of various front-end configurations while maintaining consistent architecture and training protocols. They introduce a comprehensive set of fairness metrics to quantify performance disparities, which is a significant advancement in the evaluation of audio systems. The theoretical foundation is well-articulated, linking frequency resolution to classification error, thereby providing a strong basis for their claims.
Results & evaluation
The experiments are well-designed, utilizing a diverse set of datasets across speech recognition, music analysis, and acoustic scene classification. The balanced sampling across languages and musical traditions ensures that the results are meaningful and generalizable. The statistical significance of the findings is rigorously tested, enhancing the credibility of the results. The performance gaps highlighted in the results section are compelling and underscore the need for alternative representations.
Reproducibility
The authors provide sufficient details regarding their experimental setup, including hyperparameters and dataset specifications, which facilitates reproducibility. The release of FairAudioBench as a benchmark for cross-cultural evaluation further enhances the reproducibility of their findings and allows other researchers to validate and build upon their work.
Limitations
While the study is comprehensive, it acknowledges limitations in geographic coverage, particularly the underrepresentation of African tonal languages and indigenous musical traditions. Additionally, the focus on single-axis biases without addressing intersectionality may overlook complex interactions between different forms of bias. Future work could expand on these aspects to provide a more nuanced understanding of audio processing disparities.
Broader impact
This research has significant implications for the development of inclusive audio systems that are equitable across cultural contexts. By challenging the assumptions underlying traditional psychoacoustic models, the authors advocate for a paradigm shift in audio processing that considers cultural diversity. The findings can inform the design of more effective speech recognition systems and music analysis tools that serve a global audience, ultimately contributing to a more equitable technological landscape. The main contribution of this paper is the identification and quantification of cross-cultural bias in mel-scale audio representations, alongside the introduction of alternative representations that significantly reduce performance disparities. This work is a critical step towards developing fairer audio systems, highlighting the importance of cultural considerations in machine learning applications.
Author of this paper? Send a correction.