When noise depends on context. Introducing automatic contextual audio denoising
What counts as noise is not always obvious. What we want to hear, and what we want to remove can change depending on the situation. Traffic sounds, for example, might be useful in urban monitoring, but distracting during a voice call. Yet most audio denoising systems still work from a fixed definition of what should be kept and what should be filtered out. This can be effective in many cases, but it can also mean removing sounds that matter or leaving sounds that do not fit the listening context.
This question is what motivated our latest research. Working with colleagues at Nokia and Tampere University, we introduced Automatic Contextual Audio Denoising (ACAD), a new task in audio processing that starts with a simple but important idea: can a system understand the context of an audio scene before deciding what should stay and what should be removed based on the context?
A new way to think about denoising
Traditional denoising usually assumes that the difference between the target sound and noise is defined in advance. ACAD takes a different approach. Rather than applying the same rule in every situation, it uses context to decide what is relevant. In our current study, that context comes from the acoustic scene itself. Sounds that are typical for that scene are treated as in-context, while sounds that fall outside the expected event distribution of that scene are treated as out-of-context.
In practical terms, this means the system is not simply learning to remove “everything except X”. It is learning to preserve the parts of a sound scene that fit the current context and to reduce those that do not. This makes denoising more adaptable and opens the door to systems that can make more meaningful audio decisions in real-world conditions.
How we studied the problem
To explore the idea, we created a benchmark dataset and built a deep learning method that infers context directly from the input audio, then uses that information to guide denoising. We then compared this approach with several alternatives, including a model without context inference and models using different conditioning variants.
On this benchmark, the context-aware approach outperformed the compared baselines on standard objective metrics. The results suggest that inferred context can help improve denoising when the goal is not only to suppress interference, but also to preserve sounds that makes sense in the current scene.
Why this is only the beginning
At the same time, this work also shows that contextual denoising is far from a solved problem. It is a challenging research direction, and the current benchmark should be seen as a controlled first step rather than a complete solution. The study also shows how much dataset design and evaluation choices can shape the results.
That is exactly why we believe the topic is worth pursuing. If future audio systems are going to work intelligently in complex environments, they will need to go beyond fixed definitions of signal and noise. They will need to understand relevance in a way that depends on context, use case and listening goal. ACAD is our first step towards that broader vision.
Why it matters for future audio systems
As audio technologies become more capable, expectations are also changing. People do not just want cleaner sound. They want sound processing that feels appropriate for the situation. Communication tools, intelligent devices, assistive listening systems and media applications could all benefit from methods that understand not only what is present in a signal, but also whether it belongs in the current context.
This is where contextual audio processing becomes particularly interesting. Instead of applying generic suppression rules, future systems may be able to make more nuanced decisions about which parts of an acoustic scene should be preserved and which should be reduced. That could lead to audio experiences that are not only clearer, but also more useful and more natural.
A research direction we are excited to explore
For us at Nokia, this work reflects the value of continuing to explore new questions in audio AI. Research progress often begins by looking again at familiar problems and asking whether they have been framed too narrowly. In this case, we asked whether denoising should always depend on a fixed target-versus-noise definition. With ACAD, we offer one possible alternative: context-dependent denoising.
By introducing the task, building a benchmark and presenting an initial modelling approach, we hope to encourage further progress in machine listening systems that can reason about audio more intelligently. We are excited to contribute this first step and to keep exploring where context-aware audio processing can go next.
Find out more in the paper here: https://arxiv.org/abs/2605.22262
