Far-field pickup is disrupted by room sound
Aliases: living-room noise · cocktail-party TV mic · SNR
What it is
A television microphone usually sits two or three metres from the speaker, with programme audio, a kitchen, conversation, a vacuum in between. Far-field pickup is disrupted by room sound is a claim about that physical channel: target speech arrives already covered by other sources in the room. It is not “the recogniser is not clever enough.” A living room is not a booth. Arrays and echo cancellation can lift SNR a little; they cannot subtract programme audio from the same speaker down to zero.
Why it happens
Intensity falls with distance, while the TV’s own speakers sit next to the microphone. Dialogue, title music, ads overlap speech in spectrum and create self-feedback: a large slice of what the mic hears as “the user” is what the TV just played. Echo cancellation estimates that path and subtracts; a delay change, a volume jump, or the user talking over a line leaves residue. Residue is treated as query, and title recognition drifts onto the line now playing.
A second talker, a child, a pet, an open window are a different interference: close in direction, also speech, so beamforming cannot split them and loudness wins. A near-talk mic on the remote shortens distance and dodges the TV speaker, but the remote has to reach the mouth. When it does not, the system is back on the far-field array in the chassis, with a different interference model.
Interference is not a flat noise floor. It is bursty, locked to the content, sometimes louder than the user. A design that passes on average SNR will fail in blocks during peak dialogue and ad breaks.
Where it stops holding
Headphones or a soundbar move programme audio off the chassis, so self-interference at the chassis mics drops, while in-room talkers remain. Remote near-talk or a phone as mic turns the problem into close-talk and weakens this mechanism. An empty room, daytime, search with the programme muted, is the lab default, not a family watching at night. A wake word is shorter and more energetic than a query, so it can catch at a worse SNR; the actual query sentence is longer and more fragile—wake success is not query intelligibility.
Applying it
- Duck or pause programme audio as soon as the user starts speaking, and show a clear “listening” state. Do not try to recognise over full-volume dialogue.
- Support chassis far-field and remote near-talk together; prefer near-talk when it is available, and make it visible which mic is listening.
- On failure, split “did not hear” from “heard but not in the library.” The first should prompt to raise the remote or mute; the second should go to correction. Do not loop the same “please say that again.”
- Verify in a real living-room level with three conditions: playing dialogue, kitchen noise, a second talker. Report success by condition, not an all-day average that flattens the collapse during ads.