machinehearing

Applications

• Noise reduction. VoIP, cellular etc.
• Text to speech.
• Speech to text.

Noise reduction

Motivation

• Improving intelligibility
• Reducing listener fatigue

Can be formulated in different ways

• Speech enhancement
• Noise reduction
• Source separation

Stationary sources are well served by traditional DSP methods. However complex non-stationary noise sources are not. Machine learning, and especially deep learning, has shown great promise here.

Usually done by computing time-frequeny masks and applying them via spectral substraction.