Telephony & VoIP Solutions at Scale

Traditional telephony operates in a half-duplex paradigm where audio is processed sequentially. Our architecture breaks this model entirely. When a call is established, FreeSWITCH's media bug API forks the audio stream into two independent channels: the caller's voice (read direction) and the callee/agent's voice (write direction). Each direction is encoded as raw PCM Linear16 at 8kHz or 16kHz and pushed over a persistent WebSocket connection to our streaming backend. The streaming backend (built on Python/FastAPI or Node.js) receives the caller audio in real-time, pipes it into a Speech-to-Text (STT) engine (like Google STT or Deepgram), obtains the transcription, sends it to the LLM for intent processing, receives the response text, converts it via a Text-to-Speech (TTS) engine (like ElevenLabs or Google TTS), and injects the synthesized audio back into the WebSocket stream — which is then written back into the FreeSWITCH channel. The entire round trip — from the caller finishing a sentence to the AI voice responding — is consistently below 300 milliseconds. This is achieved through aggressive buffering strategies, pre-warming TTS connections, and running inference on GPU-accelerated nodes co-located with the telephony servers.

Voice quality in VoIP is determined by three critical factors: codec selection, jitter buffer management, and network path optimization. We configure our FreeSWITCH deployments to negotiate the optimal codec per call leg. For internal LAN calls, we prefer G.722 (wideband, 16kHz) or Opus for superior clarity. For PSTN interconnects, we use G.711 μ-law/A-law to avoid unnecessary transcoding latency. For WebRTC browser clients, Opus is the unanimous choice due to its adaptive bitrate capabilities. Jitter buffers are tuned dynamically. For AI voice agent calls, we use aggressive dejitter settings (20ms packet time, 60ms buffer depth) to minimize latency at the cost of tolerating minor packet loss. For traditional business calls, we use conservative settings prioritizing audio smoothness. We also implement SRTP (Secure RTP) encryption on all call legs, with DTLS-SRTP key exchange for WebRTC endpoints and SDES for SIP trunks, ensuring end-to-end media encryption without measurable performance degradation.