Stop a voice assistant from reacting to its own speaker
If an assistant responds to words you did not say, its own spoken output may be returning through the microphone. The resulting transcript can look like an invented conversation. Compare the timing and wording before assuming every strange response comes from the language model.
Look for the echo signature
Start with a clean session and a quiet room. Ask for one short reply and remain silent while it speaks. Watch whether a new input transcript repeats part of that reply. A new user turn beginning immediately after playback is a useful clue, especially when the text resembles the assistant’s previous words.
A bad translation alone does not prove acoustic echo. Recognition errors, misunderstood instructions and prior context can also produce strange answers. Separate those possibilities by testing a short ordinary phrase first, without translation or other people speaking.
Use a harmless two-turn probe
Say: “Please say: the paper boat is green. Then wait.” Do not repeat the sentence yourself. If the app transcribes “the paper boat is green” as a new input while you stay silent, note the time and output route. End the session rather than letting the assistant keep answering itself.
Repeat with the speaker volume lower, keeping your position unchanged. Then repeat with headphones and the phone in the same place. If the duplicate input disappears with headphones, speaker-to-microphone leakage is a stronger explanation. This proposed comparison is not a measured echo-cancellation benchmark.
Make the acoustic path easier
Use moderate playback volume and avoid resting the phone in a resonant corner or directly beside a loud external speaker. A quiet furnished room is a better diagnostic environment than a reflective kitchen. Headphones reduce the amount of synthesized speech reaching the phone microphone.
If your device or installed app offers relevant microphone processing controls, test them one at a time and record the result. Do not assume a particular echo-control switch exists. Bluetooth can change input as well as output behavior, so establish the built-in route before comparing accessories.
Distinguish interruption from a feedback loop
When you deliberately speak during playback, the assistant may stop to listen. That is different from it starting a new input turn while you remain silent. A loop may grow because every generated reply becomes material for another reply. A prompt telling the model to ignore itself cannot reliably repair the acoustic capture path.
In GPT Realtime Assistant, check whether spoken output is needed for the current task. If your version allows disabling it, text replies provide a temporary way to continue without adding speaker audio to the room. Translation speech and assistant speech should be checked separately.
Report the reproducible condition
Provide support with the phrase, route, speaker volume range, device and app version, and whether headphones stopped the duplicate input. Mention whether the transcript repeats the output exactly or only roughly. These observations help distinguish audio leakage from a response-generation problem.
Do not include private recordings unless you deliberately choose to share them through the publisher’s support process. If the assistant keeps creating turns while you are silent, end the session and use text until resolved. Repeated cycles are neither useful conversation nor a safe basis for judging translation quality.
FAQ
Can a stronger prompt fix echo?
A prompt can constrain responses but cannot reliably prevent speaker audio from reaching a microphone. Test the capture and output paths.
Does improvement with headphones prove the model was fine?
It strengthens the acoustic-leakage explanation, but you should still verify the transcript and the answer for the original task.
