NoiseTag – AI -based source classification inside NorCloud

Why separating activity and non-activity noise is critical in environmental measurements.

Noise source classification in environmental monitoring

Environmental noise measurements rarely contain only the sound of the activity under study. Traffic, nearby industry, weather-related noise and other sources are often captured together with the target noise.

Identifying which events belong to the activity — and which ones do not — is one of the most time-consuming and subjective parts of noise analysis, especially in long-term or unattended monitoring projects.

NoiseTag addresses this challenge by automatically classifying noise sources during environmental monitoring. Integrated into NorCloud, NoiseTag identifies which events are relevant to the activity being studied and which are not. Irrelevant noise—sounds that are not part of the activity—can therefore be excluded from the analysis. By reducing the need for manual review, NoiseTag shortens analysis time and improves the consistency and accuracy of the results.

What is NoiseTag?

NoiseTag is an AI-based module integrated into NorCloud that automatically classifies noise events recorded during environmental monitoring. Its purpose is to distinguish between noise generated by the activity under study and noise originating from unwanted sources.

By separating relevant and unwanted noise contributions directly within the monitoring project, NoiseTag helps ensure that analysis and reporting focus on the acoustic impact of the activity itself.

How does NoiseTag work inside NorCloud?

NoiseTag integrates seamlessly into NorCloud and works directly on incoming sound events. Once enabled for a project, it analyzes detected audio events and assigns them to predefined labels.

No additional hardware is required, and setup takes only a few minutes. NoiseTag operates entirely within the NorCloud workflow, using existing measurement data and audio recordings.

How are activity and non-activity noise defined?

NoiseTag relies on a user-defined source labels focused on the sources that are relevant to the analysis. All defined source labels are considered part of the activity by default, except for those explicitly marked as non-activity by the user.

Noise sources that have not been trained or explicitly categorized are treated as unknown. By design, unknown events are considered part of the activity. This approach avoids the need to classify every possible sound and allows the analysis to focus on excluding only those sources that are known to affect the results and do not belong to the activity under study.

If a more detailed classification is required, additional source categories can be introduced and trained as needed, but this level of detail is optional and project-dependent.

How is the AI model trained?

At the beginning of a project, the AI model is trained using representative audio samples selected by the user and assigned to the appropriate source labels. In practice, only a small number of samples is sufficient to achieve reliable classification, although additional samples may be included if desired.

This project-based training allows NoiseTag to adapt quickly to the specific acoustic conditions of each site. Unlike generic pre-trained models, classification is based on sounds that actually occur within the project, ensuring that results reflect local conditions, operating patterns and source characteristics.

How confident are the classifications?

Each classified noise event is presented with a color-coded confidence score and a percentage value. These indicators show how certain the model is about each classification.

Events with low confidence, or events that do not match any trained source labels, may be labeled as unknown. Unknown events are treated as part of the activity by default and do not require immediate action unless the user chooses to further refine the source labels.

Confidence scores help users decide when manual review may be required and provide transparency in how classification results are derived.

Partial noise analysis and activity-only noise levels

NoiseTag enables NorCloud to calculate partial noise levels by excluding non-activity contributions. Depending on the project requirements, these contributions may be fully excluded, replaced by a representative background level, or addressed using percentile-based approaches, allowing the most appropriate method to be applied in each case.

Model adaptation and retraining

NoiseTag can be retrained at any point during a project to reflect changes in the source labels or classification criteria. Retraining allows the model to adapt when additional activity types are introduced or when a finer level of source categorization is required.

This ensures that classification remains relevant and accurate throughout long-term monitoring projects.

Typical use cases for NoiseTag

NoiseTag is particularly useful in projects where multiple sound sources are present and only part of the measured noise is relevant for assessment, such as:

Acoustic projects with multiple overlapping sound sources
Environmental noise monitoring influenced by non-activity noise
Compliance studies requiring activity-only noise levels
Long-term unattended monitoring with a high volume of noise events

NoiseTag and Noise compass

NoiseCompass combined with NoiseTag enables automatic noise source identification, determining both where the noise comes from and what type of source it is, even in complex environments with overlapping sources.

Reporting and data handling in NorCloud

All NoiseTag classifications and partial noise levels are available directly in NorCloud reports. Results can be included in standard reports or exported in multiple formats to create customized documentation.

Audio clips, event lists and metadata remain accessible at all times within NorCloud, and measurement data are automatically provided by instruments such as the Nor145 and uploaded to the platform as part of the normal monitoring workflow.