Testing Approaches for Adaptive Audio in Google Meet's Merged-Audio Scenarios

Most of us have experienced the awkward moment when two or more colleagues join the same online meeting from the same room. One person speaks, and a split second later the same speech is heard again through another participant's microphone. This duplicated audio can quickly make conversations distracting and difficult to follow.

Modern conferencing applications attempt to reduce these issues using technologies such as noise suppression, echo cancellation, and automatic gain control. While these algorithms perform well in many situations, they are primarily designed to remove background noise and cancel acoustic echo, where the audio played through a device's loudspeakers is re-captured by the same device's microphone. However, the scenario where multiple participants join the same meeting from the same room presents a different challenge. Instead of removing echoes produced by a device's own speakers, the application receives multiple recordings of the same speaker from nearby devices. Since each microphone captures the speech from a slightly different position, the signals arrive with small differences in delay and volume. Distinguishing between these nearly identical audio streams is considerably more complex than traditional acoustic echo cancellation.

To address this challenge, Google Meet introduced Merge Audio, an adaptive audio feature designed to allow multiple nearby devices participating in the same meeting to function as a unified audio system. Instead of treating each participant as an independent audio source, the application detects that the devices are located in the same physical space and coordinates their microphones to minimize duplicated speech and improve overall call quality.

For QA engineers, however, this raises an important question: how can we reliably verify that such an adaptive algorithm actually works? Unlike traditional features with deterministic inputs and outputs, Merge Audio depends on real-world acoustic conditions, including the distance between participants, propagation delay, and signal attenuation. Reproducing these conditions consistently in a laboratory environment is far from trivial.

This article presents the proof of concept (PoC) our team developed to establish a repeatable methodology for testing Merge Audio. By simulating realistic acoustic conditions in a controlled environment, we aimed to demonstrate that adaptive audio algorithms can be evaluated consistently and produce reliable data for further analysis.

Hardware & setup

From concept to prototype

Before assembling the laboratory environment, we first needed to answer a fundamental question: could the acoustic conditions required by Merge Audio be reproduced in a controlled laboratory environment? Simply connecting several laptops together would not be sufficient. The setup had to simulate how speech propagates through a room while remaining deterministic and repeatable.

To validate our assumptions, we designed several prototype setups and evaluated different approaches for routing and manipulating audio signals. These early designs helped us identify practical limitations and refine the final architecture before investing additional time in implementation.

Final laboratory setup

The final laboratory setup consisted of an Audio Support MacBook, two sender devices, and one receiver device (Figure 2). The Audio Support MacBook generated the required propagation delay and signal attenuation before forwarding the modified audio stream to the second sender. This allowed us to simulate two nearby participants while maintaining complete control over the acoustic parameters.

The additional sender device enabled simultaneous audio transmission from two participants, while the dedicated audio interfaces provided deterministic routing between the physical devices. RCA splitters allowed a single audio source to be distributed across multiple audio paths without altering the original signal, closely matching the requirements of our simulation.

Why latency matters

Although the Merge Audio proof of concept generated propagation delay through software, understanding the magnitude of real acoustic propagation was essential when designing the simulation. Our methodology relied on reproducing delays of only a few milliseconds. For example, sound travels approximately 1.5 metres in 4.37 ms, while a distance of 3 metres corresponds to only 8.75 ms. These values illustrate how small the timing differences between nearby participants actually are.

The methodology

Testing Merge Audio by simply placing laptops at different distances in an office would make the results difficult to reproduce. Even small changes in participant position, room acoustics, or background noise could influence the outcome. To produce consistent and comparable measurements, we chose to simulate the acoustic environment instead of relying on physical movement.

Our methodology was based on two fundamental properties of sound propagation:

• Propagation delay – sound requires a finite amount of time to travel through air.
• Signal attenuation – the received sound level decreases as the distance from the source increases.

Using standard acoustic equations, we calculated the expected propagation delay and signal attenuation for several representative distances used throughout the PoC. By reproducing these two characteristics, we were able to emulate participants positioned at different distances while keeping the laboratory environment deterministic and fully repeatable.

The calculated values used throughout the PoC are summarized in Table 1 below.

Table 1. Propagation delay was calculated using the speed of sound in air (≈343 m/s at 20 °C). Relative volume was derived using the inverse-distance attenuation model.

Scope of the simulation

Although the proposed methodology provides a repeatable way to simulate participant distance, it intentionally models only the dominant characteristics of sound propagation: propagation delay and signal attenuation. Real meeting rooms introduce many additional acoustic effects, including sound reflections, reverberation, and environmental noise, all of which influence the audio captured by nearby microphones.

Rather than attempting to reproduce every aspect of a physical meeting room, the PoC focused on creating a deterministic and repeatable testing environment. This approach provided a practical balance between realism and experimental control, enabling consistent comparison of Merge Audio behaviour across multiple test executions.

Measuring performance

A repeatable simulation is only valuable if it produces measurable results. To evaluate the effectiveness of Merge Audio, we collected objective audio-quality and network-related metrics throughout every test execution.

The collected metrics included:

• POLQA – an objective measure of perceived speech quality.
• Audio delay – end-to-end latency between transmitted and received audio.
• Network traces – verification that the observed behaviour was not influenced by network instability, packet loss, or bandwidth fluctuations.

These measurements provided quantitative data that complemented the manual validation process and helped ensure that the observed behaviour reflected the application rather than external environmental factors.

References

• Yamaha. Acoustic Echo Canceller White Paper (June 2011, PDF)
• Sengpiel Audio – Sound Level Difference Calculator
• Wikipedia – Speed of Sound

The testing process

From setup to execution

Once the laboratory environment had been prepared, the next challenge was ensuring that every test execution remained consistent. Since adaptive audio algorithms are highly sensitive to timing, even small differences in the execution order could affect the results. To minimize variability, every scenario followed the same predefined sequence.

Each test execution consisted of the following high-level steps:

• Configure the required network conditions.
• Join the conference using two sender devices and one receiver.
• Manually approve the Merge Audio prompt.
• Start synchronized audio playback and recording.
• Execute the predefined audio scenario.
• Stop all captures and collect the generated data.
• Process and validate the recorded results.

Synchronization matters

Unlike traditional functional testing, this PoC required multiple independent devices to perform coordinated actions. Audio playback, network capture, and recording had to begin at precisely controlled moments to ensure that each test represented the intended acoustic scenario. Without synchronization, even identical configurations could produce different results.

Automation vs manual validation

Automation

Automation handled the repetitive and time-sensitive parts of the testing process, including synchronized audio playback, audio recording, network capture, and post-processing of the collected data. Before each test scenario, the automation played a short preliminary audio sample designed to satisfy the conditions required for Google Meet to display the Merge Audio activation dialog. After the dialog had been manually accepted, the actual test scenario was executed automatically, ensuring consistent execution across repeated test runs.

Manual validation

Although most of the testing workflow was automated, conference creation, participant joining, and enabling the Merge Audio functionality were intentionally left as manual steps. While these interactions can be automated within automation setup, performing them manually simplified the proof of concept and provided greater visibility into the feature activation process.

Each scenario was executed seven times to improve confidence in the collected data. Once all executions within a scenario had been completed, the recorded audio, network traces, and generated metrics were manually reviewed to identify anomalies that objective measurements alone might not reveal. Unexpected bitrate fluctuations, packet loss, synchronization issues, or unusual audio artifacts could invalidate an otherwise successful test. Manual inspection therefore served as an essential quality gate before the results were included in the final analysis.

Advanced simulations: Echo and room acoustics

From distance simulation to echo investigation

Successfully simulating participant distance demonstrated that adaptive audio behaviour could be reproduced in a controlled laboratory environment. However, the proof of concept also inspired a broader question: Could the same methodology be extended to investigate real-world echo issues reported by users? Rather than developing a completely new testing environment, we explored whether the existing proof of concept could serve as the foundation for a more advanced acoustic simulation.

During the project, we investigated the Google Meet scenario in which all participants located in the same meeting room had successfully formed a Merge Audio group, while another participant joined remotely. Under certain conditions, the remote participant occasionally heard portions of their own speech returned as an audible echo. Since the root cause of the issue was unknown, our objective was not to reproduce a specific bug, but to determine whether the existing testing methodology could be extended to evaluate similar acoustic scenarios in a repeatable manner.

Building upon the original Merge Audio triggering architecture (Figure 2), we redesigned the laboratory environment to investigate this echo scenario. The new setup replaced the second sender with an additional receiver, allowing two merged participants to receive the same remote audio stream. The proposed architecture required configurable propagation delay, signal attenuation, and room-effect processing before the audio entered the conferencing application (Figure 3). Identifying a practical way to implement these processing stages became the focus of the subsequent engineering investigation.

Investigating low-latency echo simulation

Unlike the initial Merge Audio proof of concept, where propagation delay and signal attenuation were generated by dedicated audio-processing scripts, the echo investigation introduced a fundamentally different challenge. The simulated acoustic path between the loudspeaker of one merged participant and the microphone of another required delays of only a few milliseconds. Under these conditions, the latency introduced by the audio hardware itself became a critical factor.

Rather than evaluating audio quality, this stage of the investigation focused exclusively on minimizing the latency introduced by the simulation. Audio fidelity was considered a secondary objective because the primary goal was to reproduce realistic acoustic propagation delays with minimal additional latency while keeping the laboratory environment deterministic.

To implement the proposed echo simulation architecture, we investigated several approaches for generating deterministic low-latency audio delay (Figure 4). The investigation began with software-based delay generation using SoX, followed by JACK with different buffer configurations, and later expanded to professional audio solutions, including Dante hardware. The final concept proposed integrating Reaper, a Digital Audio Workstation (DAW), to combine delay simulation, signal attenuation, and room-effect processing within a single audio pipeline.

Each approach represented a different trade-off between implementation complexity, achievable latency, and hardware requirements. While software-based solutions provided a practical starting point, the evaluation demonstrated that consumer-grade USB audio interfaces introduced approximately 10–15 ms of round-trip latency—comparable to, or even greater than, the acoustic delays being simulated. Consequently, the hardware itself became the limiting factor rather than the delay-generation software.

The investigation therefore identified professional low-latency audio platforms, such as PCIe- and FPGA-based interfaces, as the most promising direction for future work. Although these solutions were beyond the scope of this proof of concept, they provide a clear path toward more accurate room-acoustic and echo simulations.

Lessons learned

The investigation demonstrated that the methodology developed for testing Merge Audio could be successfully extended toward more advanced acoustic scenarios. However, it also revealed that the challenges associated with echo simulation differ significantly from those of distance simulation.

While the original proof of concept relied on software-generated propagation delay and signal attenuation, the echo investigation introduced a new constraint: the latency of the audio hardware itself. The target propagation delays derived from representative meeting-room distances were only a few milliseconds, making the round-trip latency introduced by consumer-grade USB audio interfaces comparable to—or even greater than—the acoustic delays being reproduced.

These findings shifted the investigation from software optimization toward the capabilities of the underlying audio hardware. Although software-based solutions proved effective for evaluating Merge Audio activation, more realistic room-acoustic simulations would require professional low-latency audio platforms.

Beyond reproducing realistic meeting-room conditions, the methodology also provides a foundation for evaluating boundary and stress scenarios that are difficult to reproduce consistently in real environments. By deliberately introducing larger propagation delays or more challenging acoustic conditions, QA teams could investigate how adaptive audio algorithms behave outside their expected operating range and help identify opportunities for improving product robustness.

Future direction

Although the complete echo evaluation environment was not implemented during the scope of this proof of concept, the investigation established a clear direction for future work. The proposed architecture illustrated how deterministic distance simulation could be combined with real-time room modelling through Digital Audio Workstations (DAWs), enabling more realistic evaluation of room reflections, participant placement, and echo propagation.

Future iterations of the methodology could also benefit from professional low-latency audio hardware, such as PCIe- or FPGA-based interfaces, reducing the influence of the simulation hardware itself and allowing acoustic phenomena to be reproduced with greater accuracy.

Rather than providing a definitive solution to every adaptive audio challenge, this proof of concept established a practical and repeatable foundation for investigating increasingly complex real-world conferencing scenarios. The experience gained throughout the investigation will help guide future work on testing methodologies for next-generation adaptive audio features.

Conclusion

Testing adaptive audio features is fundamentally different from testing traditional conferencing functionality. Instead of validating deterministic inputs and outputs, QA engineers must account for physical phenomena such as sound propagation, signal attenuation, device synchronization, and room acoustics. Reproducing these conditions consistently is essential for obtaining reliable and meaningful test results.

The proof of concept presented in this article demonstrated that Google Meet Merge Audio can be evaluated using a repeatable laboratory methodology. By combining controlled audio simulation, synchronized automation, objective quality measurements, and manual validation, we established a testing approach capable of producing consistent and repeatable data for adaptive audio analysis.

Beyond validating Merge Audio activation, the investigation also demonstrated that the same methodology can serve as the foundation for more advanced research into room acoustics and echo-related scenarios. Although additional engineering challenges remain, the work established a practical framework that can be extended as adaptive audio technologies continue to evolve.

As conferencing applications become increasingly intelligent, reliable testing methodologies will be just as important as the algorithms themselves. Establishing repeatable engineering processes enables developers and QA teams to evaluate adaptive audio features with greater confidence and ultimately deliver a better communication experience for end users.