💻 computer science

Fast-SDE: Efficient Single-Microphone Sound Source Distance Estimation in Reverberant Environments

This paper introduces Fast-SDE, a lightweight, single-microphone framework that leverages a subband-based backbone to efficiently estimate sound source distance in reverberant environments, addressing the hardware and computational constraints of embodied platforms.

Original authors: Jiang Wang, Runwu Shi, Yaozhong Kang, Benjamin Yen, Takeshi Ashizawa, Kazuhiro Nakadai

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiang Wang, Runwu Shi, Yaozhong Kang, Benjamin Yen, Takeshi Ashizawa, Kazuhiro Nakadai

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 standing in a large, empty room with a single microphone in your hand. A friend is talking to you from somewhere else in the room. In a perfect, silent world, you could guess how far away they are just by how loud their voice is. But in the real world, walls bounce sound around, creating echoes that muddy the signal. Figuring out exactly how far away your friend is, using only that one microphone and amidst all those confusing echoes, is a very hard puzzle.

This paper introduces a solution called Fast-SDE. Think of it as a "smart, lightweight detective" for sound that can be installed on small robots or devices that don't have powerful computers inside them.

Here is how the paper explains it, broken down into simple concepts:

1. The Problem: The "Heavy" vs. The "Light"

Most current methods for guessing sound distance use a microphone array (a bunch of microphones spaced out). This is like trying to find a sound source with two ears; you can tell direction and distance easily by comparing what each ear hears. However, building these systems is expensive, requires precise calibration, and takes up a lot of space and power.

The authors wanted to solve this using just one microphone (monaural). The problem is that a single microphone is like a person with one ear in a noisy, echoey room. It has to guess the distance by listening to subtle "clues" hidden in the sound waves, like how the room's echo changes the sound.

Existing computer programs that do this are like heavy, slow-moving trucks. They are accurate but require massive amounts of computing power, making them impossible to run on small, battery-powered robots.

2. The Solution: The "Subband" Strategy

The authors built Fast-SDE, which is like a lightweight, agile bicycle compared to the heavy truck. It is designed to run on devices with very limited memory and processing power.

Here is the clever trick they used, explained with an analogy:

The Old Way (The Wide Net): Imagine trying to understand a song by looking at the entire sheet music at once. It's a huge, complex picture that takes a long time to analyze.
The Fast-SDE Way (The Subband Split): Instead of looking at the whole song at once, Fast-SDE cuts the sound into small slices of frequency (like separating the bass, the mid-range, and the treble).
- It splits the sound into several "subbands."
- It uses a single, shared "brain" (a shared encoder) to analyze each slice. It's like having one expert who looks at the bass, then the mid-range, then the treble, one by one, rather than hiring a whole team of experts to look at everything at once.
- This makes the math much simpler and faster.

3. How It Works (The Assembly Line)

The paper describes the process as a three-step assembly line:

The Splitter: The sound is chopped into frequency slices (subbands).
The Shared Encoder: A lightweight neural network (a type of AI) looks at each slice. It learns to spot specific patterns in the echoes that change depending on how far away the sound is. Because it uses the same "brain" for every slice, it saves a huge amount of space.
The Fuser and Predictor: The clues from all the slices are glued together. A final, simple "regression head" (a small calculator) takes these combined clues and spits out a single number: the estimated distance in meters.

4. The Results: Fast and Accurate

The authors tested this in two ways:

Simulation: They created thousands of virtual rooms with different shapes and echo levels. Fast-SDE was able to guess the distance almost as accurately as the heavy, complex methods, but it was much faster and used far fewer computer resources.
Real World: They put the system on a small robot with a single microphone and a loudspeaker. Even in a real room with real echoes, the robot could estimate the distance of the speaker.

Key Takeaways from the Tests:

Speed: The system is so efficient it can run on a tiny microcontroller (the ESP32-S3), which is the kind of chip found in simple smart home devices, not just powerful computers.
Accuracy: It made very small errors (often less than 25 centimeters off).
The "Wall" Factor: The researchers found that the system works best when the sound source and the microphone are at similar distances from the walls. If one is close to a wall and the other is far away, the echoes get confusing, and the guess becomes less accurate.

Summary

Fast-SDE is a new, super-efficient way for robots to "hear" how far away a sound is using just one microphone. By breaking the sound into small, manageable pieces and using a shared, lightweight brain to analyze them, it allows small, cheap robots to understand their spatial environment without needing a supercomputer. The authors have even made their code available for others to use.

Technical Summary: Fast-SDE

Problem Statement

Sound source distance estimation (SDE) is a critical capability for human–robot interaction, influencing speech acquisition reliability and interaction naturalness. While existing SDE methods predominantly rely on microphone arrays to exploit inter-channel cues (e.g., phase and intensity differences), these systems demand careful hardware synchronization, geometric calibration, and significant computational resources. These requirements limit their deployment on size-constrained and computationally limited embodied platforms.

Conversely, single-microphone (monaural) SDE offers a hardware-efficient alternative but remains highly challenging. Unlike arrays, a single microphone lacks explicit spatial cues and must infer distance indirectly from subtle acoustic patterns within the received waveform. In reverberant environments, these patterns are heavily obscured by room acoustics, source-content variability, background noise, and microphone characteristics, making the relationship between the observed signal and source distance difficult to model. Furthermore, existing learning-based monaural approaches often utilize heavy models with high computational costs, rendering them unsuitable for real-time applications on resource-constrained devices.

Methodology

The authors propose Fast-SDE, a lightweight, single-microphone SDE framework designed specifically for deployment on robots with limited memory and computing resources. The core innovation lies in a subband-based backbone that decomposes the frequency axis into multiple non-overlapping subbands, rather than processing the full spectrum with a wide backbone.

Network Architecture

The Fast-SDE architecture consists of four main stages:

Input Representation: The model accepts a three-channel time–frequency feature input ( $X_{in}$ ) constructed from the Short-Time Fourier Transform (STFT) of the audio signal. This includes the log-magnitude spectrum and the sine/cosine maps of the phase spectrum to preserve phase information without discontinuities.
Subband Splitting: The input feature map is decomposed into $N$ non-overlapping subbands. This design reduces computational complexity and allows the network to focus on band-specific structures, as low- and high-frequency components often exhibit different statistical characteristics.
Shared Subband Encoder: A single set of parameters (a shared encoder) processes each subband independently. This encoder utilizes lightweight residual blocks composed of pointwise and depthwise 2D convolutions. It maps each subband to a compact latent representation, learning the correlation between acoustic structure and time–frequency patterns at low computational cost.
Frequency Pooling and Fusion: The subband representations are pooled along the frequency axis and concatenated. A $1\times1$ Conv1D layer fuses these descriptors into a unified temporal feature sequence.
Regression Head: The fused sequence is converted into a distance estimate. The standard Fast-SDE variant employs a Gated Recurrent Unit (GRU) for temporal modeling, followed by temporal pooling and a Multi-Layer Perceptron (MLP). A more lightweight variant, UltraFast-SDE, replaces the GRU with a purely feed-forward design (temporal pooling + MLP) to further reduce latency.

Training Strategy

The models are trained to minimize the Mean Squared Error (MSE) between predicted and ground-truth distances. To ensure robustness, the training data is generated using the FRAM-RIR simulator across varying room configurations, microphone positions, and reverberation times (RT60).

Key Contributions

Lightweight Framework: The introduction of Fast-SDE, a monaural SDE framework optimized for embedded systems and robots with strict size and computational constraints.
Subband-Based Processing: A novel architectural design that processes frequency subbands through a shared encoder, balancing the preservation of distance-discriminative acoustic information with reduced model complexity and inference latency.
Dual Variants: The provision of two model variants: Fast-SDE (with GRU temporal modeling) for higher accuracy and UltraFast-SDE (feed-forward only) for extreme resource constraints, such as microcontrollers.
Open Source: The authors have open-sourced the code to facilitate broader research.

Experimental Results

The authors evaluated Fast-SDE through extensive simulation and real-world experiments.

Simulation Experiments

Datasets: Three datasets (Group 1, 2, and 3) were generated with increasing room diversity (1, 10, and 100 rooms) to test generalization.
Performance: Fast-SDE achieved competitive accuracy (Mean Absolute Error, MAE) compared to state-of-the-art methods like SELDNet and attention-based models. For instance, on Group 1, Fast-SDE achieved an MAE of 0.13m, matching SELDNet.
Efficiency: The proposed methods demonstrated significantly lower parameter counts and inference latency. UltraFast-SDE, with only 9.6K parameters, achieved an inference latency of 1.61ms on an NVIDIA V100 and 5.23ms on an AMD CPU. Crucially, it successfully ran on an ESP32-S3 microcontroller (4MB flash, 2.5MB RAM) with a latency of 1018ms, enabling one inference per second—a capability not demonstrated by the compared heavy models.
Generalization: As expected, estimation error increased with the diversity of acoustic environments (e.g., Group 3 MAE of 0.23m for Fast-SDE vs. 0.24m for SELDNet), highlighting the difficulty of generalizing to unseen room configurations.

Real-World Experiments

Setup: Experiments were conducted on a mobile robot (Vstone 4WDS ROVERX40A) equipped with a single microphone and LiDAR for ground-truth distance measurement.
Fine-tuning: Models pre-trained on simulation data were fine-tuned on real-world data from a specific room.
Findings: Fast-SDE achieved the lowest MAE (0.22m) among all compared methods in the real-world setting. The experiments revealed that the difference in distance between the source and the microphone to the nearest wall significantly impacts estimation accuracy (correlation coefficient of 0.77), with larger differences leading to higher errors.

Significance and Claims

The paper claims that Fast-SDE successfully addresses the trade-off between estimation accuracy and computational efficiency in monaural SDE. By avoiding the heavy computational costs of attention mechanisms and full-spectrum processing, the framework enables real-time distance estimation on hardware platforms previously considered too limited for such tasks, including embedded devices and microcontrollers.

The authors note that while the method performs well, the challenge of generalization across diverse acoustic environments remains. They modestly conclude that future work will focus on extending the framework to more challenging 3D scenarios and handling multiple sound source types. The work is supported by JST BOOST, Grant No. JPMJBS2430.