What does noise suppression mean in Zoom, Teams or on a phone?
The app removes sound it judges is not speech before anyone else hears it. Zoom offers Auto, Low, Medium and High, and at High it removes even background speech from other people (Zoom Support, read 6 October 2026). Microsoft Teams defaults to Auto: High removes all background sound that is not speech, Low keeps music audible, and Off suits high-fidelity microphones in quiet rooms (Microsoft Support, read 6 October 2026). On iPhone the same idea is called Voice Isolation, available in phone calls from iOS 16.4 (Apple Support, published 7 November 2025). The higher the level, the more sound counts as noise, including music and other people’s voices.
How does noise suppression work?
The software splits the signal into short frames and frequency bands, estimates how much of each band is noise, and turns those bands down. RNNoise, an open-source example, keeps that signal processing and adds a recurrent neural network that sets gains for 22 frequency bands, looking only 10 ms ahead so it can run on live calls (Jean-Marc Valin, RNNoise, 27 September 2017). A live suppressor has milliseconds to decide, so it guesses with little context. In browsers it is one on or off flag on the microphone track, noiseSuppression in WebRTC (MDN, last modified 28 September 2026). It resembles voice activity detection, which only decides whether speech is present, and differs from echo cancellation, which removes a known playback signal.
Does noise suppression improve speech recognition accuracy?
Often it does not. A study of single-channel speech enhancement found that the artifacts it introduces, not leftover noise, caused most of the damage to automatic speech recognition, and that mixing some original signal back in helped (Iwamoto et al., Interspeech 2022, revised 30 March 2022). Deepgram recommends skipping noise suppression entirely for pure transcription (Deepgram docs, read 6 October 2026). In pub noise below 10 dB signal-to-noise ratio, Whisper beat all 14 LibriSpeech-trained models its authors tested (Radford et al., 6 December 2022). Modern recognizers learn from noisy audio, so a denoiser can remove cues they rely on. Compare errors with suppression on and off on your own recordings.
Where does noise suppression help a voice agent?
Deepgram places most of its value in voice agents, where it cuts false barge-in events and echo, and advises starting with the platform’s echo cancellation and adding more only when it measurably helps. On a live call the payoff is fewer false interruptions, not a better transcript. Full-duplex voice AI also needs echo cancellation so the agent does not hear itself.
How do we handle noise in our own audio and voice agent work?
Our media pipeline transcribes recordings with whisper-large-v3 on a hosted inference endpoint, and the voice in our videos ships exactly as recorded: no noise gate, compressor, limiter, EQ or loudness normalization, because a noisy or quiet take is fixed at the microphone, not in the edit. For voice agents, the launch checklist we published in voice agents in production requires a test set covering accents, noise, corrections, silence and adversarial requests, and scores a pilot on action accuracy, whether the final record matches the caller’s request. That is where voice agent evaluation should judge any suppression setting.