Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
167 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Broadband DOA estimation using Convolutional neural networks trained with noise signals (1705.00919v2)

Published 2 May 2017 in cs.SD and stat.ML

Abstract: A convolution neural network (CNN) based classification method for broadband DOA estimation is proposed, where the phase component of the short-time Fourier transform coefficients of the received microphone signals are directly fed into the CNN and the features required for DOA estimation are learnt during training. Since only the phase component of the input is used, the CNN can be trained with synthesized noise signals, thereby making the preparation of the training data set easier compared to using speech signals. Through experimental evaluation, the ability of the proposed noise trained CNN framework to generalize to speech sources is demonstrated. In addition, the robustness of the system to noise, small perturbations in microphone positions, as well as its ability to adapt to different acoustic conditions is investigated using experiments with simulated and real data.

Citations (247)

Summary

  • The paper proposes a CNN-based classification approach for broadband DOA estimation using phase information from the STFT and noise signals for training.
  • It bypasses traditional feature extraction by directly processing phase maps from microphone arrays, ensuring adaptability in varying acoustic conditions.
  • Experiments demonstrate robust performance in adverse environments and superior generalization compared to conventional methods like SRP-PHAT.

Overview of "Broadband DOA Estimation Using Convolutional Neural Networks Trained with Noise Signals"

The research by Soumitro Chakrabarty and Emanuel A. P. Habets addresses the challenge of Direction of Arrival (DOA) estimation for broadband signals in adverse acoustic environments. The proposed solution leverages a Convolutional Neural Network (CNN) framework that diverges from traditional feature extraction methods, focusing solely on the phase components of Short-Time Fourier Transform (STFT) coefficients. This approach benefits from using synthesized noise signals in the training process, simplifying the construction of the training dataset and potentially enhancing generalization to broadband speech sources.

Problem Formulation and Methodology

Chakrabarty and Habets reformulate the DOA estimation challenge as a classification problem where the objective is to map inputs from microphone arrays into distinct DOA classes. The CNN-based framework bypasses explicit feature extraction; instead, it directly feeds the phase of STFT coefficients to a CNN, allowing feature learning through supervised training.

Key aspects of the methodology include:

  • Input Representation: The phase components of the STFT are structured into a matrix, or phase map, representing time-frequency information across multiple microphones.
  • CNN Architecture: The model predominantly comprises convolutional layers without pooling which process these phase maps, followed by fully connected layers to generate posterior class probabilities.
  • Training Framework: Utilizing noise signals enables the model to avoid the complexities associated with speech signals, such as silent frame detection.

Experimental Evaluation

Chakrabarty and Habets validate their approach through extensive experimentation. Notable findings include:

  • Generalization Capability: The CNN trained on noise signals successfully generalized to speech sources, maintaining robust performance across varying noise levels.
  • Acoustic Variability: When tested in acoustic conditions different from the training environment, the CNN maintained superior performance compared to conventional methods like SRP-PHAT. This highlights the model's adaptability to different room characteristics and its resilience to reverberation and noise.
  • Robustness to Microphone Perturbations: The paper demonstrates the CNN's robustness, attributed to its ability to learn through weight sharing, which enables it to manage small positional deviations in microphone arrays effectively.
  • Real-world Application: Experiments using real acoustic measurements (from existing databases) confirmed the model's adaptability, indicating potential practical deployment in various environments.

Practical and Theoretical Implications

The implications of the work are multifaceted, involving both practical deployment scenarios and future research directions:

  • Practical Deployment: This method simplifies the training phase by sidestepping dataset preparation hurdles linked to speech signals. It offers a scalable solution to deploy DOA estimation in real-time applications like teleconferencing and hands-free communication systems.
  • Future Research Directions: Areas for further exploration include adapting the model to account for multiple concurrent sound sources and testing the system’s robustness across different noise types.

The presented DOA estimation framework showcases how leveraging modern neural architectures can bridge traditional signal processing challenges, suggesting a promising trajectory for future advancements in audio and acoustics research. The streamlined training process and robust performance underline the model's viability in practical and diverse audio processing applications.