IS3 : Generic Impulsive-Stationary Sound Separation in Acoustic Scenes using Deep Filtering

Clémentine Berger, Paraskevas Stamatiadis, Roland Badeau, Slim Essid

Abstract

Acoustic scenes are typically composed of two distinct components: stationary backgrounds (e.g., wind, fan noise) and impulsive events (e.g., claps, door slams). Separating these is crucial for various audio processing tasks. This paper introduces IS³ (Impulsive–Stationary Sound Separation), a deep-learning framework that leverages a two-stage Deep Filtering architecture to surgically decouple these components. By utilizing a novel data generation pipeline that ensures clean ground truth, IS³ significantly outperforms traditional signal processing methods like HPSS and standard time-domain separation models.

Introduction

In real-world audio, "stationary" background textures and "impulsive" high-energy bursts often overlap, creating complex acoustic scenes. Traditional audio separation research has largely focused on specific domains like speech enhancement or music source separation. However, general acoustic scenes—the sounds of everyday life—require a more flexible approach.

Impulsive sounds and stationary backgrounds possess fundamentally different characteristics, necessitating distinct processing strategies. While traditional methods like Harmonic-Percussive Sound Separation (HPSS) or wavelet filtering exist, they often suffer from "leakage" or require extensive manual tuning. IS³ addresses these limitations by providing a robust, automated pre-processing tool for applications ranging from bioacoustics to professional audio mixing.

Methods: The Deep Filtering Architecture

The IS³ framework adapts the DeepFilterNet architecture, operating in the frequency domain with a two-stage approach designed for both performance and computational efficiency.

The IS³ architecture featuring a two-stage decoding process: ERB-based gain prediction and complex-valued Deep Filtering.
  1. ERB Stage: The model first predicts real-valued gains for Equivalent Rectangular Bandwidth (ERB) bands. This provides a coarse, efficient initial separation across the full frequency range.
  2. Deep Filtering (DF) Stage: For frequencies below ~6 kHz, where complex overlaps are most prevalent, the model predicts complex-valued filters. Unlike simple masks, these filters utilize neighboring time-frequency bins to capture local correlations, resulting in significantly cleaner extraction of impulsive events.

The system uses an encoder-decoder structure optimized with a combination of compressed-magnitude spectrogram loss and multi-resolution STFT loss to maintain natural audio quality.

Data Generation Pipeline

A major hurdle in training separation models is the lack of high-quality ground truth data where clean impulses are perfectly matched with clean stationary backgrounds. To overcome this, the authors developed a sophisticated data generation pipeline.

The IS³ data generation pipeline, including automated impulse removal for stationary backgrounds and random placement of isolated impulsive events.
  • Background Cleaning: Stationary backgrounds were curated from datasets like Dcase2018 and CochlScene. An automated "impulse removal" step using Gabor decomposition was applied to ensure the backgrounds were truly stationary.
  • Impulse Sourcing: Impulsive events were sourced from ESC50, Nonspeech7k, and synthetic sources. The SALT framework was employed to ensure a balanced representation of sound classes.
  • Mixture Synthesis: The pipeline generated 50 hours of training data by randomly placing impulses over cleaned stationary backgrounds at various signal-to-noise ratios (SNR).

Key Insights and Results

The researchers found that Deep Filtering is superior to traditional ratio masking for impulsive sounds. Because impulses are transient, capturing phase and local correlations via complex filtering is essential for high-fidelity reconstruction. Furthermore, the "cleanliness" of the training data proved vital; if the stationary background contains residual clicks, the model's performance degrades.

Evaluated using the Scale-Invariant Signal-to-Distortion Ratio (SI-SDR), IS³ consistently outperformed HPSS, wavelet-based methods, and Conv-TasNet. Notably, IS³ showed significantly reduced leakage, preserving silence in the impulsive track more effectively than baseline methods. Despite its performance, the model remains lightweight at approximately 2.2 million parameters, supporting near-real-time applications.

Conclusion and Future Work

IS³ represents a significant step forward in general acoustic scene decomposition. While the model achieves high performance, future work will address "mixture consistency"—ensuring separated parts sum perfectly to the original—and extending high-precision Deep Filtering to the full frequency range. Additionally, refining the handling of rhythmic, repetitive impulses (like a jackhammer) remains an area for further exploration.

Summary updated

BibTeX
@inproceedings{berger2025is,
  author = {Clémentine Berger and Paraskevas Stamatiadis and Roland Badeau and Slim Essid},
  title = {{IS3 : Generic Impulsive-Stationary Sound Separation in Acoustic Scenes using Deep Filtering}},
  booktitle = {WASPAA},
  year = {2025},
  doi = {10.1109/waspaa66052.2025.11230927},
}
Figure