Introduction
Direction-of-Arrival (DOA) estimation—the process of determining the spatial origin of a sound source—is a fundamental component in technologies ranging from noise-canceling headphones to smart home assistants. Despite its importance, current DOA models often struggle to generalize from "sterile" synthetic training environments to the complex, semantically rich acoustic scenes of the real world. Furthermore, high-performance models frequently demand computational resources that exceed the capabilities of low-power edge devices.
This paper introduces LightDOA, a highly efficient neural network designed for real-time DOA estimation. To train this model effectively, the researchers developed the Both Ears Wide Open (BEWO) dataset, leveraging Large Language Models (LLMs) to simulate realistic acoustic environments with diverse semantic contexts.
LLM-Aided Simulation Pipeline
The core innovation in data generation is a three-step pipeline that moves beyond simple noise convolution. By integrating LLMs, the researchers ensure that the training data is not only physically accurate but also semantically grounded.
- Data Preparation: Audio samples from AudioCaps are filtered to ensure single-source descriptions and high semantic similarity between audio and captions (using CLAP scores).
- Attribute Induction: GPT-4(o) is used to enrich simple captions (e.g., "she is talking") with spatial attributes such as source position, distance, and scene size.
- Audio Simulation: These attributes parameterize physical simulators like Pyroomacoustics or gpuRIR to generate spatialized stereo audio.
LightDOA Architecture
LightDOA is engineered for extreme efficiency, utilizing only ~39k parameters. The architecture focuses on the Interaural Phase Difference (IPD) as the primary spatial cue. To maintain a low computational footprint, it employs depthwise separable convolutions—a technique popularized by the MobileNet family—followed by a Gated Recurrent Unit (GRU) for temporal modeling. The DOA task is framed as a classification problem, where the 180° front-facing arc is discretized into specific angular bins.
The BEWO Dataset
The BEWO dataset, derived from AudioCaps, provides the diverse acoustic scenes necessary for robust training. It includes a wide array of non-speech sounds, such as water pouring or animals barking, situated in varied environments.
Due to the dual-microphone setup lacking Head-Related Transfer Functions (HRTFs), the dataset utilizes a front-back mapping to resolve ambiguity, folding the full 360° field into a 180° arc.
Key Results
LightDOA demonstrates that efficiency does not require sacrificing accuracy. In benchmarks against significantly larger models, LightDOA achieved:
- 57.96% accuracy at 5° resolution, outperforming the CRNN (~311k parameters) and MTL-DOA (~285M parameters).
- 85.45% accuracy at 20° resolution.
These results underscore the effectiveness of combining a lean architecture with semantically rich, LLM-aided training data.
Conclusion and Future Work
LightDOA provides a scalable solution for DOA estimation on edge devices. While current limitations include front-back ambiguity and a focus on single-source scenarios, future work will explore multi-source localization, elevation estimation, and the integration of HRTFs to further bridge the gap between simulation and reality.
Summary updated