Computational Auditory Scene Analysis

In one of the most classic references, (Bregman 1990)  [1], several listening experiments were conducted to find out how people perceived sounds. Most experiments used simple sounds such as sinusoidal tones or sysnthetic FM and noise which give good insights into basic auditory perception. On the other hand, the sound simplicity is subject to questionable value when it comes to understanding a complex sound mixture in the real world. Nevertheless, the findings led to the field called Computational Auditory Scene Analysis (CASA), coined by Brown and Cooke  [2], where these acoustic cues are used to group the sound events together into correct sources. The acoustic cues available for uses in sound source segregation are

1. Common time onset(and possbly offset)
2. Common amplitude and frequency modulations
3. Harmonicity (special to audio signal)
4. Frequency proximity (closer harmoinics fuse better than further pairs)
5. Spatial location (for more than one microphone)
6. Streaming (high level grouping, requiring cognition and music structure understanding)

• Sinusoid Modeling
• General Scene Analysis
• Phase incoherence
• Non-stationary model
• Artificial Auditory Neural Network