Voice emotion models lack a universal best fit for cash research
No speech emotion recognition model works best across all settings, a Bank of Spain review published on October 13, 2026 finds. The paper connects vocal physiology and artificial intelligence to Neurocash® research on cash perception.
From vocal cues to trust
Lucía Canalejo Guijosa, Fernando León Martínez and José María Sánchez Echave link vocal physiology, acoustic features and Speech Emotion Recognition (SER) in Occasional Paper No. 2618. Emotional and cognitive states alter breathing, vocal-fold vibration and articulation, affecting pitch, intensity, voice quality and timing.
Within Neurocash®, voice analysis offers a non-invasive complement to gaze, touch, electroencephalography and other biometric measures for studying cognitive workload, perceived security and implicit trust in banknotes.
SER involves audio selection, preprocessing, feature extraction and classification.
Model performance depends on language, emotional categories, data quality and recording conditions; no approach consistently outperforms all others.
Decades of evidence, uneven benchmarks
The narrative review assessed approximately 80–100 documents from 1980–2026, with searches conducted between October 2025 and February 2026.
It used neither PRISMA nor GRADE assessment tools and performed no meta-analysis.
Classical models suit limited datasets and offer greater interpretability; deep and self-supervised models require more computation and are harder to interpret.
Tools include audEERING, iMotions, wav2vec 2.0, HuBERT and MixedEmotions.
Cross-language transfer and naturalistic emotion capture remain unresolved.
A framework, not validation
The paper presents a conceptual basis for voice analysis in cash research, not a validated banknote-design test.
Differences in datasets and laboratory conditions limit direct comparisons between models.
The authors call for more naturalistic data, transparent evaluation and safeguards for privacy, consent and fairness.