Sound source classification for soundscape analysis using fast third-octave bands data from an urban acoustic sensor...

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Soundscape exploration requires accurate sound source characterization.
• Pre-trained audio neural networks (PANNs) excel at identifying over 500 sound sources.
• Urban noise sensors typically provide low-resolution fast third-octaves data, incompatible with PANNs' high-resolution Mel spectro-temporal input requirements.

Purpose of the Study:

• To evaluate the performance of PANNs using fast third-octaves data after transformation.
• To assess the effectiveness of a previously developed transcoder in adapting low-resolution audio data for PANNs.
• To explore the potential of this approach for urban soundscape monitoring.

Main Methods:

• Utilized a previously developed transcoder to convert fast third-octaves data to Mel spectro-temporal representations.
• Employed pre-trained audio neural networks (PANNs) with the transcoded data for sound source prediction.
• Conducted a qualitative analysis on a large-scale dataset of fast third-octaves recordings.

Main Results:

• The PANNs' performance in predicting the perceived time of presence of sound sources was not significantly degraded when using transcoded fast third-octaves data.
• The transcoder effectively adapted the low-resolution audio data for use with PANNs.
• The study confirmed the feasibility of using PANNs with readily available urban noise data.

Conclusions:

• Pre-trained audio neural networks can be effectively utilized for soundscape monitoring with lower-resolution urban noise data via a data transformation method.
• This approach maintains the integrity of sound source classification, offering new possibilities for urban acoustic environment analysis.
• The findings support the development of advanced urban soundscape monitoring systems using existing sensor networks.