Automated Anomalous Sound Detection with OpenTest Sequence Mode and AI Analysis

In motor, headphone, speaker, and complete product testing, anomalous sound is often not a simple sound pressure level issue. Many early defects appear as slight rubbing noise, whistling, buzz, anomalous noise, or changes in spectral details. Traditional electroacoustic testing and manual listening inspection are difficult to keep consistent in large-scale testing. Using TWS earbud AI listening inspection as an example, this article explains how OpenTest Sequence Mode can connect acquisition, signal analysis, AI Analysis, Pass/Fail judgment, and result output into a repeatable automated anomalous sound detection workflow.

In R&D validation and production testing for headphones, motors, and complete products, anomalous sound detection has always been a typical engineering challenge.

Taking TWS earbuds as an example, a common test approach usually includes two parts: traditional electroacoustic testing first, followed by manual listening inspection to determine whether anomalous sound exists. Traditional electroacoustic testing can cover basic metrics such as frequency response, distortion, and noise floor. Manual listening is then used to identify issues closer to real listening perception, such as buzz, rubbing noise, voice coil rubbing, air leakage, or anomalous noise floor.

In actual production environments, however, manual listening has clear limitations. Different operators may have different sensitivity and judgment criteria for anomalous sounds. The same operator may also be affected by fatigue, environment, and experience at different times. For slight, intermittent, or complex anomalous sounds, manual inspection is difficult to keep consistent over time. It also depends heavily on human resources, making it hard to scale in mass production.

OpenTest Sequence Mode and AI Analysis provide a more standardized path for this type of application. Engineers can configure product connection, audio playback, acoustic signal acquisition, data analysis, AI judgment, result saving, and report output as a standard test sequence. This helps anomalous sound detection move from subjective manual judgment toward a workflow based on standardized testing, data analysis, and Pass/Fail results.

Anomalous Sound Detection: Time and Frequency Domain Limits Are Not Always Enough

The difficulty of anomalous sound detection is that defects do not always appear as a clear single-metric violation.

In TWS earbud testing, some defective samples may have frequency-domain curves that almost overlap with good samples, making them difficult to distinguish with simple upper and lower limit lines. Some transient anomalous sounds may also fail to create stable, visible differences in time-domain waveforms. In other words, even after spectrum analysis or waveform review, engineers may still face cases where good and defective samples look similar in the data but sound different in reality.

Similar issues can also occur in motors, fans, compressors, and speakers. Bearing anomalies, slight rubbing, electromagnetic noise, loose structures, or transient noise may only appear under specific operating conditions, within specific frequency ranges, or in short time windows. If testing relies only on fixed thresholds or manual listening, early defects may be missed, while operator differences may also lead to inconsistent judgments.

Therefore, anomalous sound detection is better treated as a complete workflow: acquire sound under stable test conditions, extract time-domain, frequency-domain, or time-frequency features through signal analysis, and then use AI Analysis to support Pass/Fail judgment.

Figure 1: Anomalies undetectable by traditional electroacoustic testing

Test Setup: Building a Stable AI Listening Test Chain

For TWS earbud AI listening inspection, the test system usually needs to complete product connection, audio playback, acoustic signal acquisition, and data analysis.

During the test, the product is placed into a test fixture or shielding acoustic box. A Bluetooth adapter connects to the earbuds and controls them to play the specified audio source. The sound output from the earbuds is captured by an artificial ear or acoustic coupler, then converted by the acquisition hardware into audio data for analysis. OpenTest then performs signal analysis and AI Analysis on the acquired WAV file or test data and outputs the Pass/Fail result.

The key to this workflow is stable test conditions. For AI-based anomalous sound detection, model judgment depends on comparable input data. Therefore, the audio source, playback method, acquisition channel, sampling parameters, fixture position, test environment, and data processing workflow should remain consistent. Only when the front-end test chain is stable can the AI Analysis result be more reliable.

Sequence Mode: Configuring AI Listening as an Automated Test Task

In OpenTest, engineers can use Sequence Mode to configure anomalous sound detection as a standard test task. For TWS earbud testing, a complete AI listening sequence can include:

• Scan the product or read sample information
• Control the shielding box or test fixture
• Connect to the Bluetooth earbuds
• Control the product to play the specified audio source
• Acquire the acoustic signal
• Run spectrum analysis, anomalous sound analysis, and AI Analysis
• Output the Pass/Fail judgment
• Save test data and upload results to the MES system

With Sequence Mode, operators do not need to manually switch modules, reconnect devices, or run each analysis step one by one. Once the workflow has been validated, it can be reused in R&D validation, sample screening, or production testing, reducing differences caused by manual operation.

This is especially important for production environments. AI listening inspection should not only judge one sample correctly, but also maintain stable cycle time and consistent output during batch testing. With Sequence Mode, OpenTest connects "connection - playback - acquisition - analysis - judgment - saving" into a standardized workflow, laying the foundation for automated testing and production management.

Figure 2: Configuring an AI anomalous sound detection workflow in OpenTest Sequence Mode

AI Analysis: Detecting Anomalous Sounds Through Time-Frequency Reconstruction

In anomalous sound detection, the core value of AI Analysis is not simply replacing upper and lower limit judgment. It is to identify anomalous features in sound data that traditional time-domain or frequency-domain limit lines may not distinguish reliably.

Using the TWS earbud AI anomalous sound algorithm as an example, the algorithm can convert the original test recording into a time-frequency spectrogram. The original recording contains both normal sound features and possible anomalous features. After the model is trained using normal product data, it can reconstruct the time-frequency spectrogram: normal features can be restored more effectively, while anomalous features are weakened or removed during reconstruction.

The system can then compare the difference between the original spectrogram and the reconstructed spectrogram, and combine time-axis and frequency-axis features for judgment. For good samples, the difference between the original and reconstructed spectrograms is relatively small. For defective samples with anomalous sounds, anomalous features lead to more obvious differences. Based on these differences, AI Analysis can output a Pass/Fail judgment.

This method is suitable for anomalous sounds that are difficult to cover with traditional limit lines, such as transient noise, rubbing noise, anomalous noise floor, voice coil rubbing, air leakage, or local frequency-band anomalies. For motor anomalous sound detection, a similar approach can also be used by converting operating sound into analyzable time-frequency features to support the judgment of rubbing, bearing anomalies, periodic anomalous noise, or electromagnetic noise issues.

Figure 3: AI Analysis identifying anomalous sounds through reconstructed time-frequency spectrograms

From Manual Listening to AI Judgment: Improving Consistency and Scalability

Manual listening is intuitive, but it is difficult to keep consistent in large-scale production. Different operators may define the boundary of "anomalous" differently, and complex or slight anomalous sounds can be affected by experience and perception.

The value of AI listening inspection is that it can turn sample experience into reusable model-based judgment. Engineers can build a model by collecting good-product data, and then use the AI Analysis step to output Pass/Fail judgments in later tests. This preserves sound-feature experience from historical samples while making the judgment process more standardized.

For R&D teams, this helps accumulate data experience across different samples, structures, and defect types. For production teams, it can reduce the subjective differences of manual listening stations and improve the scalability of the test process.

AI Analysis should not be viewed as separate from engineering judgment. Engineers can still review the original audio, time-domain waveform, spectrum curve, time-frequency spectrogram, and AI output. When an anomalous sample appears, the corresponding data can also be traced for defect localization and process improvement.

Figure 4: OpenTest AI Analysis outputting Pass/Fail judgment for anomalous sound detection

Result Output: Making Anomalous Sound Judgment Traceable

In R&D validation and production testing, anomalous sound detection needs more than a simple pass or fail result. The result must also be reviewable and traceable.

With OpenTest, test data, analysis results, AI judgment results, and sample information can be managed around the same test task. For anomalous samples, engineers can further review the corresponding audio file, spectrum result, time-frequency features, or test records to locate the time range and frequency range where the anomaly occurs. Batch test results can also support sample comparison, quality tracking, and later process optimization.

This means anomalous sound detection no longer depends only on what an operator hears at the moment. Instead, it becomes a reviewable data record. For teams that need to continuously improve product acoustic quality, this traceability is often more important than a single test result.

Figure 5: OpenTest autosave

Application Scenarios: From R&D Validation to Production Screening

OpenTest Sequence Mode and AI Analysis can be applied to multiple anomalous sound detection scenarios:

• Motor anomalous sound detection: identify bearing anomalies, rubbing noise, electromagnetic noise, periodic anomalous sounds, and related issues.
• Headphone anomalous sound detection: identify buzz, rubbing noise, voice coil rubbing, air leakage, anomalous noise floor, and related issues.
• Speaker anomalous sound detection: identify distortion-like anomalous sound, rubbing, assembly issues, cavity leakage, and related issues.
• Home appliance and complete product NVH testing: identify anomalous operating sounds from fans, motors, compressors, pumps, and similar components.
• Production end-of-line testing: configure stable workflows as standard sequences to reduce manual listening differences.
• R&D prototype comparison: compare acoustic performance across different structures, materials, and assembly designs.

In these scenarios, the value of OpenTest is not only completing one acoustic test. It helps teams turn anomalous sound detection into a reusable, traceable, and scalable workflow.

The core challenge of anomalous sound detection is to consistently identify "anomalous" sounds in complex operating conditions and large sample volumes. Traditional thresholds and manual listening still have value, but in repetitive testing, batch screening, and early defect identification, teams need more standardized workflows and more efficient data analysis tools.

With OpenTest Sequence Mode, engineers can configure connection, playback, acquisition, analysis, AI judgment, and result output as a standard test workflow. With AI Analysis, the system can output Pass/Fail judgments based on sound features and help identify anomalous changes that traditional time-domain or frequency-domain methods may not distinguish consistently.

For TWS earbud, motor, speaker, and complete product testing teams, this provides a more engineering-oriented path for anomalous sound detection: acquire data through a stable test chain, use AI Analysis to support judgment, use standardized sequences to keep workflows consistent, and use test records to support later traceability and improvement.

The free version of OpenTest is available from the OpenTest official website. To learn more about OpenTest Sequence Mode, AI Analysis, or anomalous sound detection solutions, please fill out the Get in Touch form below.