Speaker
Description
Road traffic noise arises from multiple overlapping sources such as tire friction, engines, exhaust, and honking. Separating these sources is essential for assessing health impacts and developing effective mitigation strategies. Most studies report only aggregate noise levels which are insufficient for environmental impact assessment. Existing source apportionment approaches rely heavily on labelled datasets from developed countries, limiting their applicability to complex urban environments like Low-and middle-income countries (LMICs). The raw audio data used in this study encompasses non-honking vehicular noise, horn noise, and ambient environmental sounds. This study therefore focuses on separating honking and non-honking vehicular noise in heterogeneous Indian urban traffic conditions where drivers exhibit persistent, unregulated honking behaviour varying systematically with road type and congestion level. Class 1 Sound level meters (SLMs) are prohibitively expensive for LMICs, and existing approaches require both an SLM and audio recorder in the field, the SLM for dBFS to dBSPL conversion and the recorder for frequency-resolved source differentiation, imposing a compounded logistical burden, particularly consequential in LMICs. This study addresses both challenges: dBFS to dBSPL conversion using white and pink noise collocation in anechoic chamber yielding octave band wise offsets (125Hz to 10kHz). Using GMM based classification and HMM based temporal sequencing, that requires no pre-labelled data, we found that honking occupies fewer traffic frames yet contributes disproportionately higher noise energy levels with a dominant spectral peak at 4kHz. Applied across 8 routes in Delhi under varying road and congestion conditions, this framework supports traffic noise regulation and environmental assessment in resource-constrained LMIC settings.