⚡ electrical engineering

ASAP: An Azimuth-Priority Strip-Based Search Approach to Planar Microphone Array DOA Estimation in 3D

This paper proposes ASAP, an azimuth-priority strip-based search approach that enhances the efficiency and real-time performance of 3D direction-of-arrival estimation for planar microphone arrays by leveraging the higher reliability of azimuth estimates to reduce the computational burden of the SRP-PHAT method.

Original authors: Ming Huang, Shuting Xu, Leying Yang, Huanzhang Hu, Yujie Zhang, Jiang Wang, Yu Liu, Hao Zhao, He Kong

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ming Huang, Shuting Xu, Leying Yang, Huanzhang Hu, Yujie Zhang, Jiang Wang, Yu Liu, Hao Zhao, He Kong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a large, empty room with a flat, round table in the center. On this table, you have placed eight tiny microphones arranged in a circle. Your goal is to figure out exactly where a person is standing in the room just by listening to their voice. You need to know two things:

Azimuth: Which way are they facing relative to the table? (Left, right, front, back?)
Elevation: Are they standing on the floor, or are they on a balcony looking down?

This is the job of DOA (Direction-of-Arrival) estimation.

The Problem: The "Search the Whole Room" Dilemma

The standard way to solve this (called SRP-PHAT) is like trying to find a lost key in a dark room by checking every single inch of the floor, the walls, and the ceiling. It's very accurate, but it takes forever. If you have a robot or a phone with limited battery and computing power, this "check everything" method is too slow to be useful in real-time.

Furthermore, there's a catch with flat (planar) microphone arrays: they are really good at telling you left vs. right (azimuth), but they are a bit confused about up vs. down (elevation). It's like having a compass that works perfectly but a level that's a bit wobbly.

The Solution: ASAP (Azimuth-Priority Strip-Based Search)

The authors of this paper propose a new method called ASAP. Instead of checking the whole room blindly, ASAP uses a smart, two-step strategy that plays to the strengths of the flat microphone table.

Think of it like finding a specific book in a massive library:

Step 1: The "Strip Search" (Finding the Aisle)

Instead of looking at every single book on every shelf in the entire library, you first look at the aisles (the azimuth).

The Metaphor: Imagine the library is divided into horizontal strips. You quickly scan these strips to find which "aisle" the book is likely in.
How it works: ASAP focuses on the horizontal direction first because the flat microphones are experts at this. It narrows down the search to a few promising "strips" of the room. It doesn't just pick one spot; it keeps a few "hot zones" (like keeping a few candidate aisles open) just in case the first guess was slightly off.

Step 2: The "Great Circle" Refinement (Finding the Exact Shelf)

Once you know which aisle (azimuth) the book is in, you don't need to check the whole library again. You just walk down that specific aisle and look closely at the shelves (elevation).

The Metaphor: Now that you know the book is in "Aisle 4," you don't need to check "Aisle 1" or "Aisle 10." You just walk up and down Aisle 4, looking specifically at the height of the books.
How it works: The method takes the best guesses from Step 1 and draws a curved line (a "great-circle arc") between them. It then does a very detailed, one-dimensional search along this line to pinpoint the exact height (elevation) of the sound.

Why is this better?

The paper tested this method in two ways: computer simulations and real-world tests in an office with an 8-microphone circular array.

Speed: It's much faster than the old "check everything" method. In tests, it was about 14% faster than the previous best "smart" method (CFRC) and significantly faster than the full search.
Accuracy: It's also slightly more accurate. It reduced the error in guessing the location by about 4% to 5% compared to the previous best method.
Real-World Proof: They tested it with real human voices in a noisy office. The new method found the speakers faster and more accurately than the competition, even with echoes and background noise.

The Bottom Line

The authors created a "smart search" algorithm that admits, "We are really good at finding left/right, so let's lock that down first, and then we'll carefully figure out up/down."

By doing this, they made sound localization on flat microphone arrays faster and more accurate, which is crucial for robots and devices that need to listen and react instantly without needing a supercomputer to do the math. They have also made their code available for others to use.

1. Problem Statement

Direction-of-Arrival (DOA) estimation is critical for applications like robotics, teleconferencing, and human-robot interaction. The Steered Response Power with Phase Transform (SRP-PHAT) is a robust baseline method for this task. However, it faces two major challenges in 3D scenarios, particularly with planar microphone arrays:

Computational Cost: Accurate 3D estimation requires an exhaustive search over thousands of candidate directions on a dense spherical grid, which is prohibitive for real-time applications on resource-constrained embedded or mobile platforms.
Planar Array Limitations: Planar arrays (common in robotics due to structural simplicity) exhibit anisotropic reliability. They provide high-resolution azimuth estimation but suffer from poor elevation estimation (mainlobe broadening and peak flattening at high elevations). Existing methods like Coarse-to-Fine Region Contraction (CFRC) treat 3D space uniformly, failing to exploit this specific characteristic, while full-grid SRP-PHAT is too slow.

2. Methodology: The ASAP Framework

The authors propose ASAP (Azimuth-priority Strip-based Search Approach), a two-stage algorithm designed specifically for planar arrays. It decomposes the 3D search problem to leverage the higher reliability of azimuth estimation.

Stage 1: Strip-Constrained Azimuth Estimation

Instead of searching the entire 3D sphere, ASAP restricts the initial search to horizontal azimuthal strips.

Search Space: The candidate space is defined as a union of horizontal strips centered at predefined elevation angles ( $\Theta$ ).
Coarse-to-Fine Region Contraction (CFRC): Within these strips, a CFRC-style search is performed.
- The algorithm starts with a coarse spherical mesh and recursively subdivides triangles.
- At each refinement level, it retains the top- $N$ unit directions.
- Spherical Cap Filtering: Around the retained maxima, spherical caps are formed to concentrate evaluations near emerging peaks, effectively locking the azimuth angle ( $\hat{\phi}$ ) while allowing multiple elevation candidates to persist.

Stage 2: One-Dimensional Elevation Refinement

Once the azimuth is locked, the algorithm refines the elevation ( $\theta$ ) using a 1D search, avoiding a full 2D re-scan. Two refinement strategies are proposed:

Meridian-Centered (MC) Refinement:
- Fixes the azimuth at the Stage 1 estimate ( $\hat{\phi}$ ).
- Searches elevation within a narrow window $[\hat{\theta} - r, \hat{\theta} + r]$ .
- Uses local quadratic interpolation around the discrete peak to achieve sub-grid resolution.
Between-Points (BP) Refinement:
- Utilizes the top-2 unit vectors ( $u_1, u_2$ ) from Stage 1.
- Performs Spherical Linear Interpolation (SLERP) along the great-circle arc connecting these two vectors.
- Samples the arc to find the optimal elevation, leveraging the geometric relationship between the two candidates.

3. Key Contributions

Azimuth-Priority Strategy: The first method to explicitly exploit the anisotropic reliability of planar arrays by prioritizing azimuth locking before elevation refinement.
Strip-Based Search: Introduces a "strip-constrained" search mechanism that reduces the 3D search space to a manageable 2D surface (strips) and then a 1D path, significantly lowering computational complexity.
Dual Refinement Mechanisms: Proposes both MC and BP refinement strategies, with BP offering superior performance by utilizing the geometric path between high-scoring candidates.
Open Source: The authors have open-sourced the code to facilitate community adoption and further research.

4. Experimental Results

The method was validated through extensive simulations and real-world experiments using an 8-microphone Uniform Circular Array (UCA).

Simulation Results

Accuracy: The BP variant of ASAP achieved the lowest Root Mean Squared Error (RMSE) across all Signal-to-Noise Ratio (SNR) conditions (clean, 3.09 dB, and 1.5 dB noise).
- Compared to CFRC, ASAP reduced RMSE by 5.87%.
- It outperformed full-grid SRP-PHAT in accuracy while being significantly faster.
Efficiency:
- At the highest grid resolution (Level 5), ASAP reduced runtime by 13.56% compared to CFRC.
- Full-grid SRP-PHAT runtime increased drastically with resolution (e.g., ~~3987s at Level 5), whereas ASAP maintained sub-100s runtimes (~~73s).

Real-World Experiments

Setup: Conducted in a reverberant office environment ( $T_{60} \approx 0.5s$ ) with human speech from 5 speakers.
Performance:
- Accuracy: BP-ASAP achieved an RMSE of 8.83°, outperforming SRP-PHAT (8.90°) and CFRC (9.23°).
- Speed: At Level 5, ASAP (BP) took 28.58 seconds for 523 recordings, compared to 36.23 seconds for CFRC and 1556.55 seconds for full-grid SRP-PHAT.
- Improvement: Compared to CFRC, ASAP reduced runtime by 13.98% and RMSE by 4.33%.

5. Significance

Real-Time Feasibility: ASAP makes high-accuracy 3D DOA estimation feasible for embedded systems and mobile robots by drastically reducing the computational burden without sacrificing robustness.
Robustness in Far-Field: By improving elevation estimation accuracy (even marginally), the method helps prevent error accumulation in target tracking, which is critical for far-field scenarios.
Future Directions: The authors note that current parameter tuning (strip width, subdivision levels) is trial-and-error. Future work aims to integrate lightweight deep learning models for systematic parameter optimization and extend the method to multi-source and moving source scenarios.

In conclusion, ASAP represents a significant step forward in acoustic source localization for planar arrays, successfully balancing the trade-off between computational efficiency and estimation accuracy through a structured, azimuth-prioritized search strategy.