Architecting real-time autonomous voice agents and conversational call pipelines for Kurdish and multilingual environments. Combines on-device/edge Whisper STT, streaming LLM agent logic with dynamic interruption handling, and localized RegaLabs-TTS speech synthesis with sub-600ms end-to-end voice latency over WebRTC and SIP telephony.

Autonomous Kurdish Call & Voice Agents

Real-time conversational agents built for voice-first interactions, outbound/inbound telephony, and customer support with native Kurdish (Sorani & Kurmanji) and Arabic fluency.

Architecture & Real-Time Voice Pipeline

Streaming Speech-to-Text (STT): Low-latency streaming phonetic transcription tailored for regional Kurdish accents and dialects.
Turn-Taking & Interruption Engine: Real-time VAD (Voice Activity Detection) and barge-in handling so users can interrupt naturally mid-sentence.
Agentic Decision Core: Sub-second LLM reasoning with tool calling (CRM lookups, booking, database queries, and knowledge retrieval).
Acoustic Streaming TTS: Direct streaming audio generation through the RegaLabs-TTS flow model with zero buffering lag.
Telephony & WebRTC Transport: Full-duplex audio over SIP gateways, Twilio/FreeSWITCH trunks, and browser WebRTC data channels.

Key Performance Benchmarks

End-to-End Voice Latency: <600 ms from user speech endpoint to initial acoustic audio chunk.
Natural Voice Timbre: Neural acoustic voice cloning matched to native Kurdish intonation.
Multi-Turn Context Preservation: Structured memory state and channel boundaries for resilient long conversations.