The Neural Telephony Pipeline
How Noviq Voice AI connects live telecom carrier streams to local AI models with sub-second turnaround.
Voice Activity Detection
Silero VAD v5 samples 20ms audio frames continuously. Detects speech onset in <1ms and marks speech offset after 400ms of user silence.
Speech-to-Text (STT)
Faster-Whisper processes the speech buffer in non-blocking async threads with a 0.19x Real-Time Factor (RTF), generating accurate Hindi/English text.
LLM Reasoning Stream
Tokens stream at 70+ tokens/sec using our local Qwen 2.5 model. Our sentence boundary chunker dispatches clauses immediately to the speech synthesizer.
Kokoro Neural Speech
Kokoro-82M synthesizes natural 24kHz audio, converted in-memory to G.711 $\mu$-law 8kHz and streamed back through Plivo carrier WebSockets.