Why WebSockets Break Telephony Voice AI: Building a 60ms Barge-In Loop on Raw 8kHz RTP Developer Wang Guo open-sourced a telephony voice AI lab that bypasses JSON-over-WebSocket wrappers in favor of raw 8kHz RTP through FreeSWITCH, using SOXR VHQ resampling to 16kHz for SenseVoice and VAD, and a 256-frame outbound ring buffer with a 0.16ms clearAudio() hard flush to achieve a roughly 60ms barge-in loop. The design splits the pipeline into two asymmetric loops so interruption detection relies on acoustic VAD rather than waiting 200-400ms for an LLM or full-sentence transcript. Author : Wang Guo @wangguode https://dev.to/wangguode Source Lab : https://github.com/wangguode/ai-agent-voice/tree/main/01-voice-agent-rtp-bargein https://github.com/wangguode/ai-agent-voice/tree/main/01-voice-agent-rtp-bargein Stack : FreeSWITCH 8kHz Raw RTP / Pipecat Full-Duplex / SenseVoice / SOXR VHQ Resampling Most Voice AI tutorials work great over a clean laptop microphone in a quiet room using browser WebSockets. The moment you connect that same pipeline to real enterprise telephony PSTN / SIP trunks / FreeSWITCH , three brutal physical realities hit at once: 8kHz G.711 PCMU / PCMA or L16 audio, while modern acoustic models SenseVoice, Whisper, Silero VAD expect 16kHz PCM. Naïve linear interpolation introduces aliasing artifacts that tank ASR accuracy and cause false VAD triggers. In our open-source lab 01-voice-agent-rtp-bargein https://github.com/wangguode/ai-agent-voice/tree/main/01-voice-agent-rtp-bargein , we bypassed generic JSON-over-WebSocket wrappers and engineered a deterministic telephony loop: PSTN / SIP Caller 8kHz RTP, 20ms frames │ ▼ UDP Raw RTP FreeSWITCH Media Gateway │ ├──► SOXR VHQ Upsampler 8kHz ➔ 16kHz ──► SenseVoice / Acoustic VAD ~18ms │ │ │ ▼ Barge-In Triggered └──◄ 256-Frame Outbound Ring Buffer ◄──── 0.16ms clearAudio Hard Flush SOXR VHQ for 8kHz <- 16kHz conversion preserves vocal formants and prevents high-frequency aliasing from poisoning the local VAD classifier. clearAudio 0.16ms —producing immediate silence on the caller's phone. A common mistake in Voice AI is waiting for the LLM or full-sentence STT transcript to confirm "the user is interrupting" . Transcription requires context windows often 200ms–400ms . By splitting the pipeline into two asymmetric loops : All reference code, architecture diagrams, and runnable barge-in experiments are open-sourced: