### Answer
Two settings work together here: denoising mode and interruption sensitivity.
1. Pick the right denoising mode (under Realtime transcription settings)
-
Remove noise (default) filters background noise with almost no distortion to the waveform, so it has no meaningful impact on speech-to-text accuracy. It will not remove loud background speech. [(1) ](
)
-
Remove noise + background speech is more aggressive and removes both background noise and background speech. It can distort the waveform and reduce accuracy in some cases, and it has a $0.005/min surcharge. It keeps only the dominant near-mic voice, so the caller must be clearly the loudest voice on the line. If call screening is enabled, don't use this mode, because it can filter out the caller's voice.
-
No denoising passes raw audio to the speech recognition model. Choose it if short responses like "sure" or "yes" are being missed.
2. Tune Interruption Sensitivity (under Speech settings)
- The range is 0 to 1. Higher values make the agent easier to interrupt, while lower values require the caller to speak longer or say more words before interrupting.
- Lower it a step at a time (for example from 0.9 to 0.8) and test with representative background noise.
- Don't set it to 0, because at 0 the caller cannot interrupt the agent at all. Lower values reduce false interruptions but also make intentional interruptions harder, so you need to find a balance.
3. Improve the caller's audio where possible
- For web calls, use a better microphone and client-side noise reduction.
- You can also prompt the agent to ask users to speak louder so they stand out from background speech.
A community thread reports that lowering interruption sensitivity mainly helps while the agent is speaking. It did not stop the agent from treating b