Deep neural network for perceived loudness calculation

Highly accurate, faster than real-time perceived loudness analysis

In acoustics, loudness is the subjective perception of the magnitude of a sound. The loudness of a sound perceived by a human listener typically differs from the physical sound level, depending on parameters such as the sound level at different frequencies, duration, and time pattern. The difference between physical level in decibels and perceived loudness can be substantial. Existing models for predicting loudness from the physical characteristics of sound require extensive calculations, so they are computationally expensive and not able to determine loudness in real time.

Researchers at the University of Cambridge have developed a compact deep neural network (DNN) model for predicting the perceived loudness of sound from its physical waveform. Their new process has the potential to allow real-time control of perceived loudness levels in audio playback and broadcasting. With this method, the loudness of real-world sounds (including music, speech, machine and appliance sounds) can be accurately predicted in real time at low cost.

Potential applications of this technology include real time monitoring and control of audio playback/broadcasting, product design and testing, audio control for web-conferencing software and home appliances, and other applications that were previously limited by high cost and inefficiencies. We are now looking for commercial partners interested in commercialising this copyright-protected technology.

Download more information