Signal-to-noise ratio directly determines whether recognition is usable
Aliases: SNR · far-field ASR · usability threshold
What it is
Signal-to-noise ratio (SNR) is target speech energy relative to the background. Below a band, recognition is no longer “a bit worse”: the voice channel as a whole loses usability—wakes become unstable, commands do not execute, dictation is unusable. In noise the first question is whether current SNR still supports input, not which suppressor is on.
Why it happens
Acoustic models learn phoneme boundaries near the training SNR. When noise raises the floor, spectral detail in consonants is buried and formants are filled, frame posteriors flatten, and language-model bias cannot restore contrasts that are gone. Far-field use adds distance loss, so effective SNR is worse than at a handset. On-device AGC amplifies noise with speech: level looks adequate, the ratio does not improve. Usability is therefore a threshold: above it, errors are mostly substitutions; below it, deletions and garbage insertions explode and intent cannot be recovered. That is a different layer from “kinds of noise interfere differently”: here the only question is whether the energy ratio still holds the channel up.
Studying it
Step SNR under controlled noise and measure wake success, command completion, and dictation WER, finding the band where usability falls through the product gate. Use stationary broadband as a baseline, then recheck in real scenes. Distance and microphone type belong among independent variables. State whether SNR is measured on the raw waveform or after enhancement; the two must not be mixed. White noise in the lab yields an overly clean threshold; kitchen and cabin spectra move the door.
Where it stops holding
A worn close-talking microphone can raise SNR; a hands-free speaker in the same room cannot. Lip reading, keys, or touch still work at low SNR, so “the room is loud” is not “interaction is impossible”—only this acoustic channel has failed. Anechoic high-SNR scores do not transfer to a café. Blaming every failure on SNR ignores barge-in from several talkers and echo-canceller collapse, which energy ratio does not summarise.
Applying it
- Declare a minimum usable SNR for critical voice tasks and measure it in cars, kitchens, and streets, not only in an office.
- When SNR falls through the gate, demote or close the voice entry and immediately expose a non-voice way to finish; do not keep pretending the device can hear.
- Log estimated SNR with task success to check the threshold, rather than watching only mean recognition rate.