Degradation in noise is not linear; past an SNR point recognition falls off a cliff

What it is

WER versus falling SNR is not a straight slope. Across a fairly wide band errors rise gently; past a band, intent success and word accuracy drop together, and the channel falls from “still usable” into “almost random.” Degradation in noise is therefore nonlinear. A product that promises by linear extrapolation from mean SNR will fail completely on the far side of the cliff.

Why it happens

Phoneme contrasts live in spectral bumps. Until noise fills those bumps to a point, a language model can still patch with context; once they are gone, frame posteriors flatten, the language model free-associates, and insertions and wild substitutions explode. The turn corresponds to features leaving the training support, not to a fixed extra handful of errors per decibel. Enhancement can slide the cliff toward lower SNR; it rarely deletes the cliff. A usability threshold says “below this SNR do not treat voice as general-purpose input.” The cliff says that near that door, metrics will not ease across—they collapse. Telling a product to “hang on a bit” beside the cliff is meaningless.

Studying it

Sweep SNR in 2–3 dB steps, plot WER and intent success, and find the knee. Draw separate curves per noise class; the cliff moves. Plot enhancement on and off. Report the knee rather than a straight line between 0 dB and 20 dB. Field SNR estimates are noisy, so place the knee with controlled overlays and then check in the field.

Where it stops holding

Raising already-high SNR barely moves WER: that is saturation at the other end, not a cliff. Strong speakers, lexicons, and language models push the cliff down; open dictation sits higher. Hardware overload (clipping) can look like a sudden collapse whose root is distortion, not additive noise. Children and whispered speech fall off the cliff at a higher SNR.

Applying it

• Write the cliff location into the capability statement, and when estimated SNR enters the knee band, switch to a degraded mode rather than still showing “recognizing.”
• When tuning enhancement, watch how many dB the knee moved, not only how many WER points dropped in moderate noise.
• Keep demos off the comfortable SNR inside the cliff; run a full task at the scene’s real SNR.

Related

• Same group: C7.12.1 Stationary and transient noise interfere with recognition by different mechanisms · C7.12.2 A microphone array can beamform toward the talker and suppress other directions · C7.12.3 Noise suppression can raise recognition rate while costing natural voice quality
• Adjacent: C7.05 Noisy Environments · C7.03 Types of Recognition Errors
• Search: SNR cliff · nonlinear degradation · intelligibility threshold