Engineering Under 180ms Turn-Taking Latency on Indian SIP Trunks: Exotel, Vobiz, and WebSockets

The Latency Threshold for Human Conversation

Human conversational psychology dictates that when two people talk on the telephone, pauses longer than 250 milliseconds are subconsciously perceived as awkward, robotic, or disconnected.

Most off-the-shelf voice AI implementations suffer from 800ms to 1,500ms latency due to sequential HTTP API calls: audio capture ➔ speech-to-text ➔ LLM completion ➔ text-to-speech ➔ buffer playback. In Indian telecom environments where cell reception fluctuates, this latency degrades user trust immediately.

Our Real-Time Full-Duplex Architecture

BoldFlow Labs achieves sub-180ms response turn-taking on Indian telecom networks by eliminating REST bottlenecks in favor of bidirectional WebSocket pipelines:

[Indian Mobile Network (Airtel / Jio / Vi)]
       │ (G.711u / PCM Audio over SIP Trunk)
       ▼
[Exotel / Vobiz Cloud Telephony Gateway]
       │ (Bidirectional 8kHz PCM WebSocket Stream)
       ▼
[BoldFlow Voice Engine Edge Pipeline]
       ├─► Streaming VAD (Voice Activity Detection) <20ms
       ├─► Ultra-Fast Localized ASR (Hindi/English/Tamil) <65ms
       ├─► Speculative LLM Token Streaming <45ms
       └─► Neural Streaming TTS Chunk Generation <50ms
       │
       ▼ (Sub-180ms Total Turn-Taking Audio Delivery)
[Caller Hears Natural Voice Response]

Acoustic Barge-In & Interruption Handling

The hallmark of natural conversation is the ability to interrupt. If a caller says "Wait, how much is the maintenance fee?", the AI must immediately cease speaking within 50 milliseconds.

We implement real-time server-side acoustic echo cancellation (AEC) combined with fast-response VAD. The instant caller speech power exceeds background noise thresholds, outbound TTS playback is terminated mid-frame and the agent smoothly pivots to answer the interruption.