Deep learning-based direction-of-arrival estimation for multiple speech sources using a small scale array.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Accurate localization of multiple sound sources is crucial for applications like human-computer interaction and surveillance.
• Existing high-resolution direction-of-arrival (DOA) techniques often require large sensor arrays or complex computations.
• Deep neural networks (DNNs) offer a promising avenue for developing more efficient and effective DOA estimation methods.

Purpose of the Study:

• To develop a high-resolution direction-of-arrival (DOA) estimation method using deep neural networks (DNNs).
• To enable the localization of multiple speech sources with a small-scale microphone array.
• To improve upon existing high-resolution beamforming techniques in terms of accuracy and efficiency.

Main Methods:

• Feature extraction from the time-frequency spectrum, including generalized cross-correlation (GCC) coefficients and mel spectrograms.
• Training a DNN with Gaussian-distributed labels mimicking spatial spectrum characteristics.
• Predicting DOAs through peak detection on the DNN's output.

Main Results:

• The DNN-based DOA estimation method demonstrated superior performance compared to conventional high-resolution beamforming techniques in simulations.
• The proposed framework effectively localized multiple speech sources within an indoor environment using a four-element microphone array.
• Achieved high-resolution DOA estimation with a compact sensor configuration.

Conclusions:

• The proposed DNN-based approach provides an effective and high-resolution solution for multiple speech source localization.
• This method offers a significant advancement over traditional beamforming techniques, particularly for small-scale array applications.
• The framework shows strong potential for real-world deployment in various acoustic environments.