Stationary and transient noise interfere with recognition by different mechanisms

What it is

Backgrounds split roughly into stationary noise (air-con, engine, fan: statistics that change slowly) and transient noise (a slam, a tap, a laugh, bowls clinking: short pulses or bursts). They interfere with recognition differently. Stationary noise mainly raises the floor and masks weak consonants; transients blow up a few frames in a short window and are often treated as speech by VAD or they scramble the endpoint. The same “loud” does not get the same countermeasure.

Why it happens

Stationary noise is an approximately estimable power spectrum; Wiener filters, spectral subtraction, and modern masking nets can predict it from longer context. Errors tend to be ongoing substitutions and deletions, especially fricatives and word-final sounds. A transient’s spectrum is unpredictable before it hits; adaptive filters cannot catch up. The pulse leaves a block in the encoder that matches no phone; the decoder either inserts a fake word or deletes the real word it overlapped. Multi-talker babble sits between: stationary-like energy with speech structure, the easiest way to fool a “non-speech” model. Usability-threshold discussions care about total SNR; here, with the same energy ratio and different time structure, the shape of the error changes.

Studying it

Overlay the same speech with different noise classes at matched SNR and compare WER mix (substitution/insertion/deletion) and endpoint error. Keep stationary broadband, impulses, and babble in separate columns. Do not report only a score on a mixed noise set. Recheck in kitchens (transient-heavy) and cabins (stationary-heavy).

Where it stops holding

Music has quasi-stationary accompaniment and beat pulses; it will not sit in one bin. Echo and room reverberation are convolutional, not additive-stationary. Wind on a microphone is close to broadband stationary, with gust transients. Calling every kitchen failure “SNR too low” misses the transient accident in which a pot-clang cuts the endpoint.

Applying it

• Pick enhancement by the scene’s dominant noise: cabin-like stationary suppression, kitchens that also protect the endpoint from pulses.
• Split eval sets by noise class; when insertions rise, inspect transients and VAD first, when substitutions rise, inspect stationary masking.
• In transient-heavy scenes, offer short commands or a key that does not depend on a continuous speech stream, so one bang cannot destroy a whole sentence.

Related

• Same group: 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 · C7.12.4 Degradation in noise is not linear; past an SNR point recognition falls off a cliff
• Adjacent: C7.05 Noisy Environments · C7.02 Endpoint Detection
• Search: stationary noise · transient noise · babble