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 demonstrate that mel-scale features achieve 31.2% WER for tonal languages compared to 18.7% for non-tonal languages (12.5% absolute gap), and show 15.7% F1 degradation between Western and non-Western music. Alternative representations significantly reduce these disparities: ERB-scale filtering cuts disparities by 31% with only 1% computational overhead, while CQT achieves 52% reduction in music performance gaps.

1. Contributions

1. Systematic evaluation of seven audio front-ends across 11 languages, 6 musical collections, and 10 European cities
2. Demonstrating mel-scale bias: 31.2% WER for tonal vs 18.7% for non-tonal languages (12.5% gap)
3. Revealing critical frequencies: 200-500 Hz where mel resolution is insufficient for tonal languages
4. Showing alternatives work: CQT (52% music gap reduction), ERB (31% across domains with minimal overhead)
5. Releasing FairAudioBench: Benchmark for cross-cultural audio evaluation

2. Results

Performance Gaps (Figure 1)

Domain	Mel Baseline Gap	Best Alternative	Reduction
Speech (Tonal vs Non-tonal)	12.5% WER	LEAF: 8.3%	34%
Music (Western vs Non-Western)	15.7% F1	CQT: 7.6%	52%
Scenes (Europe-1 vs Europe-2)	5.6% Acc	ERB: 5.0%	11%

Comprehensive Results (Table 1)

Front-end	Speech WER/CER (%)		Music F1 (%)		Scenes Acc (%)		Overhead
	Tonal	Non-tonal	Non-West	West	Europe-1	Europe-2
mel	31.2±1.2	18.7±0.8	56.7±2.1	72.4±1.5	71.2±1.4	76.8±1.2	1.00×
ERB	26.4±1.0	17.8±0.7	62.8±2.0	73.1±1.4	72.6±1.3	77.2±1.1	1.01×
Bark	27.2±1.0	18.1±0.8	61.9±2.1	72.8±1.5	72.2±1.3	76.9±1.2	1.01×
CQT	28.8±1.1	19.2±0.9	65.3±1.9	72.9±1.4	-	-	1.15×
LEAF	25.8±0.9	17.5±0.7	62.4±2.0	73.5±1.4	72.5±1.3	77.5±1.1	1.08×
SincNet	30.8±1.1	18.5±0.8	58.3±2.1	72.5±1.5	71.4±1.3	76.9±1.2	1.06×
mel+PCEN	28.9±1.1	18.2±0.7	59.2±2.2	72.6±1.5	72.3±1.3	77.1±1.1	1.04×

Fairness Metrics

Metric	Formula	Speech	Music	Scenes
WGS	min(Acc)	68.8→74.2	56.7→65.3	71.2→72.5
Δ	max-min	12.5→8.3	15.7→7.6	5.6→5.0
ρ	min/max	0.85→0.90	0.78→0.90	0.93→0.94

3. Experimental Setup

Datasets

Speech Recognition CommonVoice v17.0

• Tonal Languages (5): Mandarin Chinese (4 tones), Vietnamese (6 tones), Thai (5 tones), Punjabi (3 tones), Cantonese (6 tones)
• Non-tonal Languages (6): English, Spanish, German, French, Italian, Dutch
• Samples: 2,000 test samples per language
• Metrics: CER for tonal, WER for non-tonal

Music Analysis

• Western Collections:
  • GTZAN (10 genres, 1000 tracks)
  • FMA-small (8 genres, 8000 tracks)
• Non-Western Collections (CompMusic):
  • Hindustani (1124 recordings, 195 ragas)
  • Carnatic (2380 recordings, 227 ragas)
  • Turkish makam (6500 recordings, 155 makams)
  • Arab-Andalusian (338 recordings, 11 mizans)
• Samples: 300 recordings per collection

Acoustic Scenes TAU Urban Acoustic Scenes 2020 Mobile

• Europe-1 (Northern): Helsinki, Stockholm, Amsterdam, London, Vienna
• Europe-2 (Southern): Barcelona, Lisbon, Paris, Lyon, Prague
• Scene Types: 10 urban acoustic environments
• Samples: 100 recordings per city

4. Implemented Front-ends In Codebase

This codebase implements 7 audio front-ends with fixed transformations that can be used as drop-in replacements:

Front-end	Type	Parameters	Description
Mel	Fixed	40 filters, 25ms window, 10ms hop	Standard mel-scale filterbank (baseline)
ERB	Fixed	32 ERB-spaced filters	Equivalent Rectangular Bandwidth scale
Bark	Fixed	24 critical bands	Psychoacoustic Bark scale
CQT	Fixed	84 bins (7 octaves × 12 bins)	Constant-Q Transform for music
Mel+PCEN	Fixed	Mel + per-channel normalization	Adaptive gain normalization
LEAF	Learnable	64 learnable Gabor filters	Data-driven frequency allocation, adapts to task
SincNet	Learnable	64 learnable sinc filters	Learnable bandpass filters, sinc-based

5. Quick Start

git clone https://github.com/shivam-MBZUAI/cross-cultural-mel-bias.git
cd cross-cultural-mel-bias

# Install dependencies
pip install -r requirements.txt

# For HuggingFace datasets (required for speech data)
pip install huggingface_hub
huggingface-cli login  # Login with your HF token

Dataset Preparation

Assuming the data files are already downloaded and are present inside data/ folder.

Preprocessing

### Preprocessing

# Process specific domains
python preprocess_datasets.py --data_dir /path/to/data --output_dir processed_data --domain speech
python preprocess_datasets.py --data_dir /path/to/data --output_dir processed_data --domain music
python preprocess_datasets.py --data_dir /path/to/data --output_dir processed_data --domain scenes

# Parameters:
#   --data_dir: Directory containing raw data with speech/, music/, scenes/ subdirectories
#   --output_dir: Output directory for processed evaluation sets
#   --domain: Which domain to process [speech|music|scenes]

# This creates balanced evaluation sets:
#   - Speech: Max 2,000 samples per language (11 languages)
#   - Music: Max 300 samples per tradition (6 traditions)
#   - Scenes: Max 100 samples per region (2 regions)

Running Experiments

# Run complete evaluation pipeline - Single run
python frontends_eval.py

# Multi-run with significance testing
python frontends_eval.py multi

# This will:
# 1. Load your processed audio data
# 2. Evaluate 7 implemented front-ends on all 3 tasks  
# 3. Calculate fairness metrics (WGS, Gap, DI)

5. FairAudioBench

We introduce FairAudioBench, the first comprehensive benchmark for evaluating cross-cultural bias in audio systems present in file 'preprocess_datasets.py' and 'frontends_eval.py':

Components

• Curated Datasets: Balanced evaluation splits across 11 languages, 8 musical traditions, 10 cities ('preprocess_datasets.py')
• Evaluation Suite: Automated computation of fairness metrics (WGS, Δ, ρ) with statistical significance ('frontends_eval.py')
• Reproducible Pipeline: Complete evaluation in single script ('frontends_eval.py')

Citation

If you use this code or our FairAudioBench dataset in your research, please cite our ICASSP 2026 paper:

@inproceedings{melbias2026,
  title={Cross-Cultural Bias in Mel-Scale Representations: Evidence and Alternatives from Speech and Music},
  author={Shivam Chauhan and Ajay Pundhir},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2026},
  organization={IEEE}
}