### Answer The issue you're experiencing—where the AI detects "Bob" twice and self-interrupts—is likely caused by your aggressive noise cancellation mode combined with fast transcription settings. Here are specific adjustments to try: 1. Adjust Denoising Mode Your current setting "Remove noise & background speech" is the most aggressive mode and may distort the waveform, potentially causing transcription issues like duplicate detections [(1) ]() . Try switching to "Remove noise" (the default) , which removes background noise with nearly no distortion to the waveform and has no meaningful impact on speech-to-text accuracy . 2. Consider Transcription Mode You're using "Fast" transcription mode. If the environment is quiet and transcription completeness is the priority, consider "No denoising" mode, which disables all audio preprocessing and passes raw audio directly to the ASR model . This can help if you're experiencing issues with missing short responses or degraded accuracy . 3. Fine-Tune Interruption Sensitivity You've already increased interruption sensitivity to 0.85, which is good. The interruption sensitivity setting helps the agent reduce false interruptions from background speech and noise . However, this is a trade-off—higher settings make the agent more resilient to background speech but also hinder its ability to be interrupted . Recommended Settings to Try: - **Denoising mode**: Switch from "Remove noise & background speech" to "Remove noise" - **Interruption Sensitivity**: Keep at 0.85 or experiment slightly higher (0.9) if the duplicate detection persists - **Wait before speak**: Your increase from 0 to 0.6 is good—this gives the ASR more time to finalize the transcription before the AI responds Why This Happens: The duplicate "Bob" detection suggests the ASR is pick

ing up echo, processing artifacts from aggressive noise cancellation, or the turn-taking model is detecting speech activity twice [(2) ]() [(3) ]() . The "Remove noise & background speech" mode can distort the waveform and reduce speech-to-text accuracy in some cases [(1) ]() .

​ To learn more, see the sources I considered: [[1] Handle background speech & noise](https://docs.retellai.com/build/handle-background-noise) [[2] VAD vs. Turn-Taking Endpoints in Conversational AI | Retell AI](https://www.retellai.com/blog/vad-vs-turn-taking-end-point-in-conversational-ai) [[3] Speech Processing](https://www.retellai.com/glossary/speech-processing) ​

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.

### Answer Since the previous settings made the self-interruption worse, try these alternative configurations: 1. Switch to "No Denoising" Mode If the background noise level is not particularly high, set the denoising mode to No Denoising [(1) ]() . This preserves more of the raw audio signal and can significantly improve ASR accuracy for brief utterances when the environment is relatively quiet [(2) ]() . The "No denoising" mode disables all audio preprocessing and passes the raw audio signal directly to the ASR model [(1) ]() . 2. Switch Transcription Mode to "Accurate" Change your transcription mode from "Fast" to "Accurate" [(2) ]() . Sometimes transcription quality can be impacted if the sentence was cut off (the transcription spits out the finalized sentence before it should) . The accurate mode can help prevent premature sentence finalization that might be causing the duplicate detection. 3. Lower Interruption Sensitivity Further Try setting interruption sensitivity to 0.7 or lower [(1) ]() [(3) ]() . The interruption sensitivity setting helps reduce false interruptions from background speech and noise [(1) ]() . Lower values mean it will take longer/more words for the user to interrupt the agent [(4) ]() . This should help prevent the AI from treating a single "Bob" as two separate speech events. 4. Increase Responsiveness Setting Lower the responsiveness value to **0.7 or

0.8** [(3) ]() [(5) ]() . Lower responsiveness means the agent will wait longer before responding [(3) ]() [(4) ]() . This gives more time for the transcription to stabilize before the AI starts generating a response. Recommended Test Configuration: - **Denoising mode**: No denoising - **Transcription mode**: Accurate - **Interruption Sensitivity**: 0.7 - **Responsiveness**: 0.7-0.8 - **Wait before speak**: Keep at 0.6 - **Voice speed**: Keep at 1.05 - **Dynamic voice speed**: Keep ON The combination of no denoising with accurate transcription mode should reduce waveform distortion while giving the ASR more time to correctly identify single utterances [(1) ]() [(2) ]() .

​ To learn more, see the sources I considered: [[1] Handle background speech & noise](https://docs.retellai.com/build/handle-background-noise) [[2] Increase transcription accuracy](https://docs.retellai.com/reliability/wrong-transcript) [[3] Step 2: Configure the basic settings](https://docs.retellai.com/build/single-multi-prompt/configure-basic-settings) [[4] Update Voice Agent](https://docs.retellai.com/api-references/update-agent) [[5] Step 1: Configure global settings](https://docs.retellai.com/build/conversation-flow/global-setting) ​

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.