Measure translation latency on smart glasses before buying

Time needed about 30 minutes including setup, five runs, and one retest

The steps

1. Choose and position the test locationPick a quiet room or area and stand or sit where you will normally converse.You should see: The location has controlled background noise and you can repeat identical deliveries from the speaker.
2. Place the recorder to capture voice and displayPosition a second phone or camera so it records the speaker’s voice clearly and also shows the glasses' display area if testing subtitles.You should see: Recorder frame includes the speaker and the glasses display area; audio waveform is clear on playback.
3. Open the translation app and select languagesLaunch the translation app on the glasses or companion phone, pick the source and target languages, and set subtitles or spoken output as required.You should see: The app shows a ready or listening state and the selected language pair is visible.
4. Prepare the scripted phrase and the speakerGive the speaker a short natural sentence and rehearse a consistent start so each run begins the same way.You should see: Speaker is ready and starts the sentence consistently on each take.
5. Record five clean runsStart the recorder, have the speaker say the scripted sentence, watch for the subtitle or translated audio, and repeat until you have five clean takes.You should see: You have five recordings each containing the speaker’s line and the glasses' subtitle visible or translated audio audible.
6. Mark timestamps for each runOpen each recording, note the timestamp of the speaker’s first audible syllable and the timestamp when the subtitle first appears or translated audio begins, and enter both into your spreadsheet.You should see: Spreadsheet lists two timestamps per run and a computed latency difference for each run.
7. Calculate mean and check for outliersCompute the arithmetic mean of the five latency values and optionally the median; flag any outliers and add notes on conditions.You should see: Spreadsheet shows mean latency, median if computed, and any flagged outliers with notes.
8. Compare the result to your thresholdDecide whether the average meets your pre-set conversational threshold for subtitles or spoken output and mark pass or fail in your notes.You should see: A clear pass/fail decision is recorded with mode and location noted.
9. Run isolation tests when results are slowRepeat the test with offline/on-device mode enabled, with phone tethering off, and on a different network to see which condition reduces latency most.You should see: You have additional latency numbers showing differences between offline/local, phone-tethered, and Wi‑Fi or cellular conditions.

Quick, repeatable test that gives a single number

Run a short, repeatable script five times and average the measured delays. That single number tells you whether subtitles or spoken translations arrive fast enough for back-and-forth conversation. Test with the exact app, settings and companion pairing you plan to use in public.

• Measure from the speaker’s first audible syllable to the subtitle becoming visible or translated audio starting.
• Repeat five identical scripted phrases and average the results.

Which latency components affect face-to-face chat

End-to-end latency is the sum of three parts: speech-to-text (transcription), text-to-display (rendering subtitles), and speech-to-speech (synthesis and playback) when audio output is used. For subtitles you measure the first two; for spoken translation include synthesis time.

• Subtitles: first audible syllable → subtitle visible.
• Spoken translation: first audible syllable → translated audio starts.

A 2–5 minute step-by-step test you can run now

You need a scripted sentence, a recorder that captures both the speaker and the glasses display (or audio), and a stopwatch or waveform view. Record five clean takes, discard runs with interruptions, and keep app settings and mic placement identical across runs.

• Script: a short, natural sentence with a clear first syllable such as “Where is the train station?”
• Recorder: position a second phone so it records the speaker’s voice and the glasses display if testing subtitles.

How to measure and report the latency

Open each recorder file, note two timestamps per run — speaker’s first audible syllable and the moment the subtitle appears or translated audio begins — subtract to get latency per run, then calculate the arithmetic mean of five runs. Record device, app, language pair and mode alongside the numbers.

• Use an audio waveform to find the first syllable timestamp and a video frame to mark subtitle appearance.
• Store runs in a spreadsheet with columns for run1…run5, average, and notes.

Practical thresholds for conversation

Set a pass/fail threshold before testing based on how fast you need replies. As a rule of thumb: under roughly 350 milliseconds feels natural for subtitles; 350–800 milliseconds is noticeable but usable; above 800 milliseconds breaks the conversational rhythm. Spoken translation tolerates slightly more delay.

• Pass: average under your conversational threshold (example: ~350 ms for subtitles).
• Borderline: 350–800 ms — expect interruptions and confirmations.
• Fail: over 800 ms — choose a different app or device for quick exchanges.

Quick isolation checks to find the bottleneck

To tell whether the delay is network, companion routing or mic/display, run three isolation checks: offline/on-device mode in airplane mode, paired-phone (tethered) mode with mobile data, and a Wi‑Fi mode. Compare averages to see where the time is added.

• Mic placement: test with the mic very near the speaker and then at normal distance to reveal gain issues.
• Continuous vs push-to-talk: push-to-talk adds a fixed startup delay; continuous removes that kickoff time.
• Language pair: rare pairs may route to cloud services and add latency.

Device and app routing you must check

Different glasses and apps process audio in different places. Some apps run models on-device; others forward audio to a paired phone or to cloud servers. Check the app listing for offline or on-device notes and test the glasses-native app where available.

• If a companion phone handles processing, latency includes glasses → phone → server → phone → glasses hops.
• Test both the companion app and a glasses-native app version where available; results often differ.
• See our app and device pages for device-specific notes: /apps/mentra-translation, /devices/ray-ban-meta, /devices/xreal-air-2-ultra, /devices/apple-vision-pro.

Common failure patterns reviewers miss

Reviewers often test once in one environment and report accuracy without repeatable latency numbers. The typical street failure is an early subtitle followed by a corrected overwrite; that pattern signals server-side reprocessing or verification delays rather than a display problem.

• Pattern: early subtitle then correction — cause: incremental local transcription followed by server confirmation.
• Pattern: translation only appears after the speaker pauses — cause: app waits for end-of-speech to improve accuracy; not suitable for quick exchanges.
• Fix reviewers miss: compare five identical runs across network conditions.

Troubleshooting steps to lower latency

If your average exceeds the threshold, change one variable at a time and retest. Start with mic placement, enable offline/on-device models, switch network routing, then try a different app or the glasses-native version. If none of that helps, choose a different device-app pair.

• Move the mic closer or use a lapel mic and re-run the test.
• Enable offline or on-device mode and re-run to avoid network round trips.
• Try the glasses-native app rather than a phone companion and compare results.

Next steps after you have numbers

If the average meets your threshold, retest in the real locations where you’ll use the glasses. If it fails, record the device-app pair and test alternatives using the same spreadsheet so comparisons are apples to apples.

• Retest in at least two likely real-world locations before committing to a workflow.
• Use our translation directories to pick alternatives and retest systematically: /uses/translation.

Where this usually goes wrong

• Subtitles appear quickly then are corrected a second laterApp shows incremental local transcription then overwrites with a server-confirmed correctionEnable offline or on-device mode if available, or use an app that presents only the final text to avoid disruptive corrections.
• Latency spikes when the phone is in your pocketCompanion app routes audio through the phone or a Bluetooth path, adding buffering and routing delaysMove the phone out of the pocket, use the glasses’ built-in mic if possible, or use a lapel mic to bypass the phone's low-gain mic.
• Latency improves dramatically in airplane modeOnline/cloud processing or an unstable network is adding extra round-tripsUse an offline model for the language pair if available or move to a stable local Wi‑Fi network; otherwise choose an app that supports on-device processing.
• Translations are slow only for a rare language pairBackend cascades or server-side fallbacks add processing steps for uncommon languagesTest a common language pair to measure baseline, and choose a service that documents on-device support for the rare language if you need it.

The checklist

• Choose and position the test location
• Place the recorder to capture voice and display
• Open the translation app and select languages
• Prepare the scripted phrase and the speaker
• Record five clean runs
• Mark timestamps for each run
• Calculate mean and check for outliers
• Compare the result to your threshold
• Run isolation tests when results are slow

Frequently asked questions