TEST DIMENSION / SOUND & TURN-TAKING
Test restaurant voice AI under noise and timing pressure
Answer in briefUse a quiet baseline and then change one acoustic or timing condition at a time. A useful result explains what was audible, how the guest and system took turns, and whether the guest could recover when the connection or conversation became difficult.
Loman Reference · Editorially reviewed November 9, 2026.
Make noise conditions reproducible
Describe the actual setup: phone model or handset, speaker mode, distance, location and background source. “Restaurant noise” is too vague if the next evaluator cannot recreate it. If sound level was measured, record the device, position and measurement method. Otherwise use a clear qualitative label and do not attach an invented decibel figure.
Keep background content separate from the intended speaker. The system should not treat a nearby conversation as the guest’s instruction. Do not use real guest recordings or copyrighted audio without the required rights and consent. A consenting test group can create a controlled background condition.
Separate four kinds of delay
- Response delay: when the guest finishes a meaningful turn to the system beginning a relevant response.
- Interruption handling: whether the system stops or yields when the guest starts correcting it.
- Pause tolerance: whether a guest can think or consult someone without being cut off.
- Recovery delay: time spent repeating or repairing after a misunderstanding or connection interruption.
Avoid a single speed score
Fast speech can hide poor understanding. A system that answers immediately but forces a full restart may feel worse than one that asks a short clarification. Report timing alongside the number of repair turns and the final interaction outcome.
For human timing, define exactly which points the reviewer uses in the recording. Use the same rule on every case and mark timing as unavailable when the recording does not support it. Do not claim laboratory precision from a manual estimate. The evidence log accepts an optional reviewer-entered response delay and labels it accordingly.
Simulate connection trouble safely
For an interrupted-connection condition, use an approved test setup and document how the interruption was created. A deliberate pause is not a dropped connection. Do not interfere with public phone infrastructure or run hazardous physical tests. The goal is to observe whether the system acknowledges uncertainty and offers a useful recovery path.
Retest the quiet baseline after making configuration changes. If both quiet and noisy cases regress, the problem may not be noise-specific. Keep every failed and blocked case in your records so it cannot disappear through selective retesting.
Customer evidence for your voice test
Restaurant reports put timing in context: shorter phone handling can free staff for guests, while simultaneous ticket flow changes the load reaching the kitchen. Test both the conversation and its handoff.
Results and operating context from Loman’s published customer case studies.
Vittorio’s Pizza
- Up to 5 min → roughly 65 sec
- phone handling time per call or orderReported before/after comparison after about six months with Loman.
- 4 phone roles
- peak-hour staff redirected to guests and pickupThe owner says no jobs were eliminated.
Use Vittorio’s approximate five-minute-to-65-second comparison to define your own timing measure. Record when the task starts and finishes, then repeat it with background noise instead of assuming a shorter call is always a successful call.
Read the Vittorio’s Pizza case study ↗Watch Vittorio’s Pizza on YouTube ↗
Euro Gyro
- 5 tickets
- seen arriving together in ToastAn owner-observed example of concurrent order flow.
- No manual re-entry
- phone orders flow directly into ToastThe workflow described in Loman’s Euro Gyro case study.
Euro Gyro describes a busy phone environment and five tickets arriving together in Toast. During overlapping test calls, review the conversation and the resulting ticket separately so noise or timing failures stay visible.
Use a paired observation, not a noise label alone
Write the same intended request twice: once on a quiet call and once with a documented background source. Record whether each attempt required a repeat, whether a correction interrupted the response, and whether the final intent was understood. Do not score background conditions you never ran.
For Loman, identify the purchased voice and transfer scope in its current plan descriptions. Then verify the behavior on the actual phone path. A list of features cannot tell you the response delay in your environment.
Published November 8, 2026 · Substantively reviewed November 9, 2026